Jobs was not talking about hospitals, clinical workflow, artificial intelligence or digital health. He was talking about where innovation must begin, not with technology, but with customer experience.

His point was simple and demanding. You cannot start with the technology and then figure out where to sell it. You have to start with the experience you want people to have, then work backward to the technology, systems and decisions required to make that experience possible.

That lesson belongs at the center of health innovation.

Too many promising companies enter the health sector leading with sophisticated platforms, powerful algorithms, elegant data architecture or novel science. Those strengths matter. However, they are not where adoption begins. Adoption begins with the people expected to use, believe in, approve, pay for, or benefit from the solution.

For health-sector entrepreneurs, the starting question cannot be, “What can our technology do?” It has to be, “What experience are we trying to create for the patient, clinician, researcher, administrator or institution we seek to serve?”

This is becoming increasingly visible as digital tools converge with therapeutics and clinical trials become more dependent on digital solutions. The question is no longer only whether the technology works. It is whether the experience works for patients, clinicians, researchers and institutions.

That is where entrepreneurial friction in health begins: not as an obstacle to creativity, but as a test of whether innovation has been shaped by the needs of the people who must use, have confidence and adopt it or by the capabilities of the technology itself.

From Fragmentation to Friction

For years, I have written about fragmentation throughout health. Long before fragmentation became a common buzzword at conferences or board meetings, it was evident that disconnected systems, competing incentives and isolated decision-making were creating unnecessary barriers for patients and clinicians alike. Fragmentation described the architecture of the problem.

Increasingly, I believe friction better describes the human experience of that architecture.

Fragmentation explains why organizations struggle to work together. Friction explains what physicians experience when they document the same information repeatedly, what nurses experience when technology disrupts rather than supports their workflow, what patients encounter when they navigate disconnected systems and what entrepreneurs discover when promising innovations stall within institutional bureaucracy.

Health professionals know this concern well from years of implementing electronic medical records. Too often, technology introduced to organize care has added clicks, documentation burden, and screen time, reminding innovators that adoption depends not only on what a system can do, but also on what it asks clinicians to absorb.

Every unnecessary approval, incompatible technology platform, duplicate workflow, unclear responsibility and poorly communicated decision creates resistance. None of those obstacles improve patient care. Each one slows the movement of innovation from discovery to implementation.

Health does not suffer from a shortage of remarkable ideas. Every week brings advances in artificial intelligence, biotechnology, precision medicine, diagnostics and digital health. Many of these innovations demonstrate meaningful improvements in clinical outcomes. Far fewer become part of everyday practice because the institutional friction surrounding implementation often receives less attention than the science itself.

Communication is a Key Implementation Strategy

Many founders in health start-ups are rightly fluent in science, engineering, data and clinical logic. That expertise is essential. The risk is that the human pathway to use receives less attention: the relationships, explanations and confidence-building that help patients, clinicians, administrators, payers and institutions understand how a new solution fits into their world. Without that connection, even strong ideas can meet resistance that looks like reluctance but often reflects an avoidable gap in understanding.

One misconception continues to undermine otherwise promising innovation. Communication is often viewed as beginning only after the product is complete. Marketing launches the announcement. Public relations introduces and positions innovation. Internal communications explain the rollout. That sequence misunderstands the purpose and impact of communication.

Communication is not simply how organizations describe innovation. Communication helps institutions understand change, reduce uncertainty and build the confidence required for adoption. It belongs alongside engineering, clinical research, workflow design and implementation planning from the earliest stages of development.

Consider a company that develops an artificial intelligence platform capable of reducing radiology turnaround times while maintaining strong clinical accuracy. The evidence is compelling. Independent validation supports the findings. Investors celebrate the technology’s potential.

Implementation nevertheless slows because department leaders worry about governance, radiologists question liability, information technology teams raise cybersecurity concerns and administrators remain uncertain about workflow integration. None of those questions challenge the supporting science. Each reflects uncertainty that could have been anticipated and addressed much earlier through supporting evidence and communication.

Consider another example. A digital platform helps people living with diabetes remain engaged between office visits, improving adherence and strengthening patient self-management. Physicians initially hesitate because they worry the technology will dramatically increase after-hours patient messages. Once the implementation team demonstrates automated triage, clearly defined clinical responsibilities and realistic workflow expectations, enthusiasm begins to replace skepticism. The technology itself remains unchanged. Understanding changes, interest grows and institutional friction begins to ease.

Communication does not replace implementation. It is part of the implementation. It turns complexity into shared understanding, aligns the people who must approve, use, pay for change, and reduces the friction that market fragmentation creates. Without communication, even beneficial innovation can remain trapped between promise and practice.

Designing Innovation for the Real World Experience

Jobs’ lesson should not be reduced to a technology slogan. It is honed and relentless discipline. Start with the experience and work backward. Keep asking whether each decision brings the user closer to value or pushes the organization deeper into layers of complexity. As Stephen R. Covey advised, “begin with the end in mind.” In health innovation, that end is not the technology itself. It is the experience, confidence and value created for the people expected to embrace and engage.

Health entrepreneurs should apply that relentless discipline within their own organizations by encouraging healthy debate among engineers, clinicians, patients, operational leaders and communicators. Diverse perspectives almost always produce stronger solutions because they test assumptions before the market, hospital, physician, or patient is forced to do so.

Equal attention should be devoted to eliminating the destructive friction that appears once innovation enters health institutions. Entrepreneurs should ask how many additional clicks a physician must complete, how many approvals a hospital must obtain, how easily the innovation integrates with existing systems, and whether every stakeholder understands not only what the innovation accomplishes but also how it improves everyday practice.

That is why one experienced health innovation champion, Levi Shapiro, founder and curator of mHealth Israel, a community of more than 20,000 health entrepreneurs, frames the challenge this way: “Clinical results and physician enthusiasm are table stakes. To overcome the ‘death by PILOT’ trap, the technology should integrate seamlessly into existing workflows, harmonize with operational and security requirements and demonstrate measurable ROI with minimal oversight. The technologies that scale are usually the ones that make adoption feel manageable, not disruptive.”

The future of health innovation will not be defined solely by better algorithms, more sophisticated diagnostics or increasingly powerful therapeutics. Success will belong to organizations that recognize implementation as a discipline requiring leadership, operational design, communication and empathy. Scientific excellence opens the door. Institutional readiness determines whether anyone walks through it.

Some friction strengthens thinking and encourages excellence. Other friction creates delay, confusion and unnecessary resistance. Recognizing the difference may become one of the most important responsibilities facing health entrepreneurs, institutional leaders and communicators alike.

Health does not face an innovation deficit. It faces an implementation deficit, made worse when communication is treated as an afterthought. Reducing the friction between what technology can do and what people need to experience may prove to be the next great breakthrough.

The post The Friction Between Innovation and Experience appeared first on Medika Life.

But this year, because of a new Medicare program, Curry has traveled a little more often.

In February, during one trip, he was told unexpectedly that he needed preapproval for the procedure. Then he went again a month or so later to get the injection, for a total of 10 hours on the road. His clinic wanted him to come in a third time, which they had never asked of him before. That appointment was “just to fill out a piece of paper to tell them how you feel again,” Curry said, so he hasn’t gone.

In January, Oklahoma became one of six states to begin a pilot program testing the use of preapprovals in traditional Medicare, the federal health insurance program for people 65 and older or with disabilities. Medicare had previously eschewed the practice — also known as prior authorization — which requires patients or someone on their medical team to seek insurance approval before proceeding with certain procedures, tests, and prescriptions.

Epidurals like Curry’s are among 13 medical services subject to the new program because the Trump administration says they’re prone to fraud or misuse. Powered by artificial intelligence, the program — called the Wasteful and Inappropriate Service Reduction Model, or WISeR — is intended to save the federal government money and protect patients from potentially unsafe or unneeded care.

Yet early reviews from Oklahoma and the other pilot states — Arizona, New Jersey, Ohio, Texas, and Washington — suggest WISeR’s rollout has not been smooth. Patients, doctors, and other healthcare professionals who spoke with KFF Health News say the effort has created confusion, errors, long wait times, and stress. Some described the rollout as “horrendous” and say people enrolled in Medicare in the pilot states are now getting ensnared in the same red tape as those with private insurance.

One key concern is that it all happened too hastily. WISeR was announced in June 2025 and launched in mid-January.

That was “quicker than normal” for the federal government, said Todd Baker, who recently stepped down as CEO of the Ohio State Medical Association. Doctors “just sort of had to figure it out,” added Jeb Shepard, director of policy at the Washington State Medical Association.

Government contractors have also acknowledged the rapid pace. “We’ve had an aggressive rollout from the time of being notified to going live,” said Jeremy Friese, CEO of Humata Health, the vendor for Oklahoma. Tech executives servicing other states have said they were still adding features to their products in the spring.

Abe Sutton, director of the Center for Medicare and Medicaid Innovation, which is administering the program, didn’t comment on the rollout schedule. But he said in a statement that the goal of these reforms is to ensure that prior authorization is efficient, fast, and streamlined.

“The model aims to reduce inappropriate care without delaying appropriate care,” he said.

Mehmet Oz, the leader of the Centers for Medicare & Medicaid Services, told NewsNation in December that they were “rolling out some prior authorization on abused practices.”

“The purpose of these is not to deny care,” Oz continued. “It’s to make sure you get the care you need and deserve, not the care some unscrupulous doctor wants to use on you.”

Medicare has struggled in recent years with suspected fraud associated with particular services. The Department of Health and Human Services’ inspector general warned in September that the program’s spending on skin substitutes, for example, had surged nearly 700% over two years, raising “major concerns about fraud, waste, and abuse.” Skin substitutes are among the 13 therapies currently subject to review under WISeR.

The program also imposes prior authorization requirements for kyphoplasty, a surgery for spinal fractures, which a report by the Medicare Payment Advisory Commission flagged as overused.

Sutton acknowledged, however, that “the percentage of providers committing waste, fraud, and abuse is small.”

Consumers and clinicians largely detest prior authorization. Even as federal health officials test the process for Medicare, the Trump administration is trying to scale it back for those with private insurance. According to a KFF poll conducted in January, 69% of insured adults consider prior authorization a burden for care.

Through WISeR, doctors and their staff log in to online portals to submit medical records that justify the procedures. Using artificial intelligence, the systems quickly approve applications that meet the program’s criteria, Friese, Humata’s chief executive, told KFF Health News. He said there is an “immediate yes” in 88% of cases for which clinical data supports an approval.

CMS has touted the process as one in which decisions are returned within 72 hours. After that, clinicians receive a “universal tracking number,” which allows them to schedule the procedure and get paid. In practice, however, participants say the process is anything but easy.

The University of Washington’s medical system alone had nearly 100 patients waiting earlier this year for epidural injections due to WISeR-related delays, according to an April report from the office of U.S. Sen. Maria Cantwell (D-Wash.) that drew on hospital association data. “Now, patients are subject to delays or denials which did not exist prior to the WISeR Model,” the report said.

Curry, the Oklahoma cattle farmer, said he might go to Kansas for future treatments to avoid the approval process. Dorota Gribbin, a New Jersey-based physical medicine and rehabilitation physician, said that by the time authorization came for one of her patients who needed a back pain procedure, the patient had gone to the hospital for more expensive care.

Jennifer Valle, a precertification and insurance supervisor at Clinical Radiology of Oklahoma, said when it comes to kyphoplasties, there has been a lot of “nitpicking” from reviewers. Other times, information her practice provides to CMS gets overlooked, she said, and reviewers ask for imaging that’s already in the file.

Claims with no problems are supposed to be paid within 15 days, said James Webb, a musculoskeletal radiologist in Tulsa, Oklahoma, who has also been frustrated by the prior approval and reimbursement process for kyphoplasties. “Six- to eight-week delays is what we’ve been seeing,” he said.

“It’s been horrendous,” said Jerry Sobel, a Phoenix-area pain management doctor. “Right from the beginning, there seemed to be no organization.” Sobel said that as of May, he hadn’t gotten paid by Medicare for nine epidurals.

“We continuously monitor operations and work closely with stakeholders to address questions and improve the provider experience,” said Sundar Subramanian, the CEO of Zyter, which has the contract for Arizona.

During an April webinar, another Zyter executive acknowledged a large backlog in payments stretching to January. Those backlogs “are currently being resolved,” Medicare’s Sutton said, without providing further detail.

When asked about other issues — including what doctors suspect are AI-driven errors — Medicare’s Sutton said the agency appreciates “feedback on provider experience.” It will be used “to help providers better understand WISeR processes,” he said.

Although CMS vendors say humans make the final decisions on approvals, doctors and their staffs believe artificial intelligence is playing a large role in the process and that denials are sometimes the result of AI hallucinations that garble or make up information.

One Arizona doctor, who wasn’t authorized by his practice to speak, recalled a denial saying his patient wasn’t eligible for procedures in the thoracic region, or mid-back. The patient needed an injection to the neck. Webb, the Oklahoma radiologist, documented four times that a patient lacked numbness, and yet his WISeR application was still denied, citing numbness, which, in the reviewer’s interpretation, would rule out the spinal surgery procedure.

Friese, Humata’s CEO, said he hasn’t heard about any AI hallucinations.

The process is also raising government costs. With more rejections, more appeals are being filed with Medicare’s administrative contractors. The government pays the contractors to handle the appeals, and Medicare’s Sutton acknowledged that the agency has “accounted for potential changes in the volume of Medicare appeals because of the WISeR program and its associated costs.”

Eighty-four percent of commercial insurers already use AI tools, according to a survey released in 2025 by the National Association of Insurance Commissioners, though they have consistently said AI isn’t used to deny prior authorization requests.

Its use in Medicare risks introducing friction and frustration into the program — and piling costs onto its beneficiaries. Prior authorization saves money for insurers partly by making patients pay a price in wait times and inconvenience, said Miranda Yaver, a University of Pittsburgh health policy researcher studying the technique.

“People will end up getting ensnared in a lot of red tape, having to be on hold, and getting rerouted,” she said. She often wonders whether prior authorization simply shifts costs to patients and doctors, rather than saving them.

Some doctors involved in Medicare’s prior authorization experiment believe it will inevitably expand beyond a few services officials in Washington consider fraud-prone.

“Everybody knows that if this pilot project works, it will be prior auth for basically all procedures,” said Mary Clarke, a family practice physician in Stillwater, Oklahoma. “If they can show that they can save money, then that’s going to be extrapolated and rolled out to other procedures and multiple other things in other states.”

When asked whether CMS is considering expansion of its prior authorization pilot, Sutton said in his statement that there are “currently no changes” considered for the list of services subject to the WISeR program, “but CMS continues to assess whether any changes are warranted.”

The post Medicare’s AI Push Snarls Patients and Doctors in Errors and Delays appeared first on Medika Life.

The body remains, but the self doesn’t.

We understand something in those moments that we rarely say plainly. And perhaps, it’s time we put this idea front and center. Cognition isn’t merely something a person has, it’s something a person is.

Day after day, we become ourselves through the act of thinking. From the complex to the trivial, we traverse a reality that bumps and bruises us into personhood. And that friction isn’t an obstacle to identity, it’s how identity forms.

Aristotle understood this long before neuroscience provided a name for it. Character isn’t something we possess. It is something we create. What we think shapes what we do. What we do, repeatedly, shapes who we become. Which is why the question of artificial intelligence, at least to me, isn’t primarily a question about productivity or efficiency.

Of course, AI doesn’t arrive as a threat, it arrives as a relief. And that’s what makes it so insidious. There’s no cognitive check engine light to warn you. There’s just the comfort of a swift and almost effortless answer. The friction that used to shape you simply didn’t happen. Do that enough times and something changes, not dramatically, but in the way that habits shift things. Gradually, then all at once.

Technology has always extended human capability. The wheel extended our legs. Writing extended memory. The calculator extended arithmetic. But AI is different in kind, and not merely degree. It reaches into cognition itself, into the territory where “we” live—into the domain of judgment, understanding, and idenity. A calculator doesn’t threaten to do your becoming for you.

The neuroscientist Michael Merzenich is well-known for the mechanism that we today call neuroplasticity. Simply put, neural connections are strengthen when used and weakened when not. The brain adapts continuously to the demands placed upon it. This isn’t a lofty metaphor but measurable biology. The brain you exercise is not the brain you don’t.

But cognitive surrender isn’t a neutral act. Every decision handed off to AI are small withdrawals from the account of the self. Of course, handing the process over to a machine provides certain efficiency or even relief, but you step away from the mechanism through which you author, well, you.

There is a phrase, adapted from the Upanishads, that I alluded to earlier: as you think, so you act. As you act, so you become. This is doing more than describing habit. It is describing identity formation. We are not simply what we know. We are, in part, what we have struggled to understand.

The answers may still sound like you. What fills the space is not.

That’s the abyss. Not a dramatic fall, but a quiet retreat from the very process that makes a person a person.

I wrote about the Borrowed Mind as a possibility. Now, I think it’s worth asking, with some regularity, whether it has become a habit.

John Nosta is the author of the best seller:  The Borrow Mind—Reclaiming Human Thought in the Age of AI.

The post AI and the Cognitive Abyss appeared first on Medika Life.

The Exhibition Floor as a Market Signal

However, after several days walking the exhibition floor and listening to discussions across multiple stages, another story emerged. The most interesting development at HLTH Europe was not the continued rise of AI. It was the growing number of companies applying artificial intelligence to solve very specific challenges faced by researchers, physicians, health systems and patients.

What appears on the stages and exhibition floor at HLTH often reflects where the market sees opportunity. Conferences do not create trends. They reveal them. HLTH Europe brought together more than 400 speakers, some 350 sponsors and approximately 5,000 participants from across the global health ecosystem. Artificial intelligence was not simply one topic among many. The conference featured a dedicated AI @ HLTH Zone, AI-focused exhibitors and numerous sessions exploring implementation, governance, clinical applications and operational adoption.

The prominence of AI across both the agenda and exhibition hall was revealing. Conference organizers dedicate space and programming to topics that matter to attendees, investors and sponsors. The visibility of AI at HLTH Europe suggested that health-specific applications of artificial intelligence have moved beyond emerging interest and are now a significant market focus.

That shift matters because health has always demanded more than technological capability. New tools must operate within environments where privacy, safety, accountability and trust are essential. Researchers are looking for ways to accelerate discovery. Physicians want to reduce administrative burdens that consume valuable time. Health systems seek efficiencies that improve operations without compromising quality. Increasingly, innovators are designing AI solutions around those specific needs.

That reality helps explain why many of the most compelling AI companies at HLTH Europe are building solutions specifically for health rather than adapting tools designed for other industries.

As Sophie Taylor-Roberts, managing partner and FINN Partners UK Health Group Lead, shared: “A mistake in healthcare carries a human cost: it can literally mean life or death. That’s why healthcare needs bespoke AI models, tools and solutions that allow for diverse patient populations, differing clinical guidelines, funding and regulatory structures.”

She added, “As with all aspects of health, one size doesn’t fit all. AI must be treated like a highly specialized medical instrument, built to respect national sovereignty, multilingual patient care, and absolute data privacy.”

Health-Specific AI Moves from Possibility to Practice

The trend was visible throughout the exhibition hall, where companies focused on clinical research, physician workflow, diagnostics, patient engagement, digital safety and operational efficiency demonstrated how specialized AI is rapidly becoming a category of its own.

The trend was visible throughout the exhibition hall, where companies focused on clinical research, physician workflow, diagnostics, patient engagement, digital safety and operational efficiency demonstrated how specialized AI is rapidly becoming a category of its own. Their growth reflects a broader shift occurring across the health sector as organizations seek tools designed for specific scientific, clinical and operational challenges.

Gabriele Ricci, Chief Data & Technology Officer at Takeda, captured that evolution when discussing AI’s growing role across the research and development continuum. “AI is transforming the future of healthcare by accelerating every stage of the R&D value chain through purpose-built capabilities tailored to specific scientific and clinical challenges,” he said.

His emphasis on purpose-built capabilities mirrors what was visible throughout HLTH Europe. The conversation is no longer centered exclusively on artificial intelligence as a technology platform. Increasingly, attention is turning toward how specialized applications can address distinct needs across research, clinical care and health operations.

Among the companies reflecting this shift was Briya, whose AI-powered platform helps researchers interact with complex data through conversational interfaces. Rather than requiring users to navigate multiple databases, coding environments and analytical tools, the platform seeks to simplify the path from question to insight.

David Lazerson, Briya’s co-founder and chief executive officer, believes many organizations misunderstand where the greatest challenge in AI adoption resides.

“Many people assume AI adoption is about choosing the right model,” he said. “In reality, the model is only a small part of the solution. The hard part is everything around it: security, governance, data harmonization, domain expertise, and the methodology required to produce trustworthy outcomes.”

His observation reflects a reality becoming increasingly evident throughout the health sector. Access to powerful AI models is expanding rapidly, shifting competitive advantage toward organizations that can generate reliable outcomes within specific health environments. That reality helps explain the growing number of exhibitors focused on narrowly defined use cases rather than general-purpose AI.

A similar perspective emerged from conversations with Keith Grimes, MD, Chief Innovation Officer at Curistica. A physician who spent 24 years in primary care, Grimes approaches artificial intelligence through the lens of risk management, governance and patient safety.

“Physicians have always governed risk,” he explained. “We do it instinctively for doctors, drugs and devices. Digital is just the fourth D, and the discipline is much the same, but it is the one we were never trained for, so the commitment to ‘do no harm’ runs ahead of the know-how.”

His comments address one of the most significant challenges facing health organizations today. Many leaders recognize the promise of AI, yet remain uncertain about implementation, oversight and accountability, particularly in smaller physician practices and community-based care settings.

Dr. Grimes emphasizes that smaller organizations should not view those limitations as barriers.

“Small practices are the cornerstone of primary care, but they cannot out-resource a hospital trust, and it does not need to,” he said. “Good governance scales down, and the same standards that protect a large organization can be borrowed rather than rebuilt.”

“We give whoever is responsible for AI and digital safety both the platform and the people,” Dr. Grimes said. “Power tools that guide them, whatever their experience, with clinical safety experts behind the software.”

Taken together, the perspectives of Dr. Grimes and Lazerson point to the emergence of a new category of innovation. The most promising health AI companies are not focused exclusively on algorithms. They are creating environments that combine technology, expertise and governance to solve specific high-friction problems.

The Future Belongs to Reliable Outcomes

For smaller organizations, this evolution may prove particularly significant. Historically, adopting advanced technology often required substantial investment, specialized technical talent and complex integration efforts. Many health organizations lacked the resources to pursue those initiatives.

Lazerson believes that model is changing. “That’s why we’re seeing the emergence of a new layer of domain-specific AI,” he said. “Instead of every organization hiring AI engineers and building custom infrastructure, they can access a complete, purpose-built environment as a service.”

The implications extend far beyond research organizations. Physician practices, community health providers, home health agencies and emerging life science companies increasingly have access to capabilities that previously required significant internal resources.

“For smaller organizations in particular, it’s a no-brainer,” Lazerson added. “They can start generating value immediately without complex integrations, dedicated AI teams, or having to solve privacy, security, and compliance challenges on their own.”

Throughout HLTH Europe, companies focused on clinical research, workflow automation, diagnostics, care coordination and patient engagement demonstrated how artificial intelligence is becoming increasingly specialized. Rather than attempting to transform every aspect of health simultaneously, they are concentrating on areas where measurable value can be achieved quickly and responsibly.

That focus on practical outcomes may ultimately become the defining characteristic of the next generation of health innovation.

Dr. Grimes summarized the principle succinctly. “Safety is not a box-ticking exercise; it works when everyone knows the part they play,” he said. “The advantage is not scale, it is fit.”

Walking through HLTH Europe, I was reminded that innovation rarely advances through a single breakthrough. More often, progress emerges through focused efforts to solve meaningful problems. The companies attracting attention were helping researchers move faster, supporting clinicians facing administrative burdens and enabling organizations to adopt new capabilities with greater confidence.

Perhaps among the more important lessons from HLTH Europe. The future of AI in health will not be defined solely by the largest platforms. It will be shaped by innovators who combine technology, expertise, and specificity to deliver reliable outcomes. As Lazerson observed, “The future won’t belong to organizations with the biggest models. It will belong to those who can turn AI into reliable outcomes.”

Judging by what appeared across the stages and exhibition floor in Amsterdam, that future is taking shape.

The post At HLTH Europe, the Most Important AI Story Was Happening Beyond the Headlines appeared first on Medika Life.

But today’s advanced digital machines differ greatly from early simple computers. The digital world aims not only to copy human memory but also to implement the structure of neurons and the complex mechanisms of human neurophysiology so that machines do not have to receive data from outside every moment and can hold the information needed to perform a task in an instant.

So far, the design of advanced digital machines seems to have worked well. Where is the problem? Why, as the digital industry advances, does the human–machine interaction still break down? Why does AI remain distant from the real world of human users, when these users—clients, patients, and service recipients—must trust the technology, share information with it, and receive information from it?

As a psychologist in the learning sciences, I believe the problem is the digital industry’s lack of attention to “human deep learning.”

Where does human deep learning come from?

Until now, attempts have focused on defining deep learning for machines, and designers claim that relying on neuroscience finishes the job. But if neuroscience were sufficient, we would see fewer challenges today. The problem is that data scientists often forget that the ultimate goal is the human being—and humans are the most complex creatures. Neuroscience is only a small part of the knowledge about the mind and human development. To design human-like machines (for example, in medicine, therapy, or rehabilitation), we must explore a much broader and more diverse range of dimensions of user learning.

I call this “human deep learning.”

Although the term “deep learning” may not appear explicitly in some learning and psychology literature, its ideas are present throughout rich scholarship in the learning sciences. Cognitivist approaches in learning science—which view learning as meaningful, durable understanding or as deep change in thinking—are aligned with deep learning. These approaches argue that when a human deeply learns a belief or concept, the change endures and transfers to different but similar situations. This kind of learning is purposeful and therefore shows up in learners’ behavior and performance.

Cognitive and learning theories that speak to the idea of deep learning include Gestalt theory, Piaget, Vygotsky, Bandura (cognitive-behavioral approaches), and others.

What should we do?

If we want to design digital machines’ deep learning with these perspectives in mind, we must focus much more on behavioral sciences and learning psychology. Then the problems of mutual understanding and interaction between machines and humans will become more manageable.

As noted earlier, AI layers have been built heavily on data science, and designers claim to mimic neuroscience, but human sciences are not limited to neuroscience. If interaction between humans and machines is to occur, the processes of meaningful human learning—seen through learning science—must be discovered and used to guide machine design.

To what extent have designers taken steps in this direction?

Conclusion: Redefining AI layers

It is time to reconsider the layers of artificial intelligence. In common models, AI foundations rest mainly on data, algorithms, machine learning, neural networks, and deep learning. But if the ultimate purpose of this technology is to serve and effectively interact with humans, we must add another foundational layer: the learning sciences.

In two recent papers I wrote for the HTLH Europe 2026 conference and published in MedikaLife , I addressed the importance of the learning sciences. The first, “Human-Centered AI in Digital Health: Why Learning Sciences Matter,” discussed why learning sciences matter in digital health, a key application area for AI. The second, “Operationalizing Learning Sciences for Human-Centered AI in Digital Health,” explained some practical principles for applying learning sciences in digital health.

In this article I tried to highlight the central connection between learning sciences and AI—what I call “deep learning.” For that reason I reshaped the common pyramids that describe AI layers according to my perspective and the arguments I presented in my papers, and I proposed a new pyramid.

Learning sciences are not just an academic field; they are the foundation for understanding how humans learn, decide, change behavior, and grow cognitively. At the core of this foundation are cognitive psychology, behavioral sciences, motivation, social learning, self-regulation, and meaningful learning.

This set is what I call “human deep learning.”

In this framework, the future evolution of AI (and all AI-linked domains such as digital health) should be modeled by a pyramid in which learning sciences and human deep learning are not a side layer but the foundation of the whole structure. If AI is to work effectively in medicine, education, mental health, and other human-centered fields, it cannot rely only on data and inspiration from neuroscience. The next generation of AI must learn that humans and their learning processes are central and foundational.

The next generation of AI must learn how humans learn.

References

Ferdosipour, A. (2026). How can instruction and learning lead to deep learning in the learner? The answer from Gestalt cognitive psychologists. The Learning Guild (Accepted/In Press, July 2026 publication).

Ferdosipour, A. (2026). Why AI needs Vygotsky: The case for AI-based intentional friction. The Learning Guild.

Ferdosipour, A. (2026). Human-centered AI in digital health: Why learning sciences matter. Medika Life. https://medika.life/human-centered-ai-in-digital-health-why-learning-sciences-matter/

Ferdosipour, A. (2026). Operationalizing learning sciences for human-centered AI in digital health. Medika Life. https://medika.life/operationalizing-learning-sciences-for-human-centered-ai-in-digital-health/

Ferdosipour, A. (2026). Why biological learning demands the friction we seek to delete? Medika Life.

Ferdosipour, A. (2026). The shift from pure modernity to human-centered modernity. Medika Life.

LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553), 436–444.

Piaget, J. (1950). The psychology of the child. Basic Books.

Vygotsky, L. S. (1978). Mind in society: The development of higher psychological processes. Harvard University Press.

Bandura, A. (1989). Social cognitive theory. Annual Review of Psychology, 40, 1–25.

Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning: A textbook. MIT Press.

The post Machine Deep Learning or Deep Learning of Humans? Which is Correct: “Machine Deep Learning” or “Deep Learning of Humans”? appeared first on Medika Life.

It goes without saying that artificial intelligence has digitalized everything these days, including the healthcare sector—ranging from mental health chatbots to health assessment and monitoring tools. While these tools are impressive in terms of quality and speed, many users may abandon them after initial use. Alternatively, there may be a lack of sufficient trust in user data privacy, and this inability to capture consumer trust can leave product developers disheartened. These are just a small fraction of the challenges and issues that the digital health sector faces today.

These concerns are not merely theoretical assumptions. Recent digital health industry reports show that sustaining user engagement and retention remains one of the most critical challenges in this field. Despite heavy investments in developing technical and AI capabilities, many digital health startups face high drop-off rates and declining user engagement after the first few weeks of use. Recent reports from Rock Health and Deloitte have also shown that trust, user experience, and user-perceived value are among the most critical factors determining the success or failure of digital health solutions.

Let us dissect the core challenge a bit more deeply.

The reality is that the ultimate success of digital health tools does not depend solely on their technical prowess, even though the most precise mathematical calculations are designed and implemented by elite engineering teams to build these tools. Rather, their ultimate success depends on the quality of the user’s “cognitive,” “motivational,” and “behavioral” experience. In other words, the core issue is not simply “what the AI knows” and how fast it delivers it to users; the golden nugget is “how the human interacts with the AI.”

In the previous article or part one, the importance of “learning sciences” in developing “human-centered AI” in digital health was discussed. I highlighted the crucial point that the missing link in AI technology, including digital health, is the absence of a vital foundation known as the learning sciences. The present article is an operational continuation of that discussion, attempting to demonstrate how the learning sciences can serve as a framework for designing cognitive and behavioral experiences in AI-driven digital health tools.

In this article, I offer recommendations that are more operational in nature for manufacturers and designers of AI tools within the digital health industry.

Learning Sciences as the Foundation for AI Design in Digital Health

The learning sciences and psychology of learning consist of a body of findings and theories regarding how a relatively permanent change occurs within an organism or learner. These changes depend on various internal and external factors. Furthermore, this change involves the learner’s cognitive, behavioral, motivational, and physiological dimensions.

With this simple description, it becomes clear that the learning sciences are not limited strictly to educational environments. Because they study the processes of cognition, attention, motivation, mental engagement, feedback, self-regulation, and behavior change in human-environment interaction, the learning sciences are vital wherever learning, interaction, action and reaction, or behavioral continuity are involved. They aid us in understanding the behaviors, motivations, cognitions, and perceptions of learners.

Especially in the era of AI, the science and psychology of learning demand deeper immersion and greater precision in constructing AI tools. This urgency arises from growing concerns that these tools may not be human-centered, neglecting the existential dimensions of the human being as the primary user.

One of the most important areas of AI application is digital health. The learning sciences can serve as a necessary prerequisite in AI design, acting as both an interpreter and a facilitator.

In what ways are they a prerequisite for AI and digital health?

They are a prerequisite because digital health tools can themselves be viewed as environments for learning and behavior change—environments where users are constantly interpreting information, making decisions, regulating behavior, and building trust. Many current digital health tools, despite their technical complexity, are not designed based on the cognitive, behavioral, motivational, and transactional complexities of human beings.

Some users have reported that the explanations provided by AI systems were ambiguous, complex, or even confusing to them. This finding aligns with the World Health Organization (WHO) report on the ethics and governance of artificial intelligence for health. The report emphasizes that explainability is only valuable when it is understandable to the end-user, as overly complex explanations can themselves become a factor in reducing trust and increasing confusion.

In certain studies, users have stated that they cannot comprehend the system’s decision-making logic and are forced to simply trust or distrust its output blindly. In many cases, the user eventually abandons these tools after a while—a reaction that a learner might similarly display in an AI-driven educational environment!

These and similar problems demonstrate that providing information clearly and orderly is not enough on its own. The interaction must be meaningful and comprehensible within the user’s cognitive dimension; the tool must understand the user’s behavior and reinforce their motivations. As a result, continuous, purposeful interaction and trust will be fostered. It is precisely through purposeful interaction and trust that the tool becomes useful and works in service of the consumer (or the learner).

Designing Tools Aligned with the User’s Cognitive Dimension

As previously stated, the human being is a creature of interwoven, complex dimensions. According to psychological and especially learning theories, a major part of the learning process occurs mentally within the dimension of cognition. Therefore, understanding the formula of learning and its cognitive dimension is an essential blueprint for designing digital tools.

For instance, in the human learning process, an unfamiliar and unknown topic transitions into a familiar one through distinct stages. This process can typically occur via mental stimulation and environmental support. If this learning is deep and meaningful (rather than superficial or based on rote, parrot-like memorization), it can be recalled for a long time, and the likelihood of forgetting is minimized.

The science of learning encompasses cognitive theories that emphasize concepts such as perception, meaningfulness, the integrated whole, problem-solving, scaffolding, and similar ideas. If designers implement these abstract concepts operationally, they can achieve practical results, including solving the following issues:

1. Information Overload: One of the most significant challenges in digital health tools is information overload. For example, in some studies conducted on chronic disease monitoring platforms, users reported that a high volume of simultaneous notifications, charts, and recommendations led to mental fatigue and a decreased willingness to continue usage. Researchers describe this phenomenon as a type of Cognitive Overload, which can even degrade the quality of health decision-making. Users often interact with these tools under conditions of stress, anxiety, or mental exhaustion. In such states, presenting a massive volume of data, alerts, or advice simultaneously can induce cognitive fatigue, confusion, and reduced decision-making quality.
2. Complex Explanations: In some research, users of AI-driven health systems reported that overly complex explanations did not increase their trust; instead, they heightened anxiety, hesitation, and mental strain. These findings demonstrate that successful cognitive design does not mean providing more information, but rather reducing mental strain and facilitating user comprehension.
3. Sustaining Attention: Maintaining users’ attention and mental engagement is another major challenge in digital health. Many health applications are abandoned by users after a short period. Reviews published in the mHealth sector show that a significant portion of health app users severely reduce or entirely stop their interaction within the first three months. This indicates that initially acquiring a user and retaining their cognitive engagement for continuous use are two completely different challenges. Part of this issue stems from the interaction experience becoming repetitive, impersonal, and cognitively tedious. The missing goal here is the personalization of the process!
4. Gradual Adaptation: On the other hand, altering a user’s attitude and perception is typically a gradual process, not an instantaneous one. Yet, some digital health tools deliver a large volume of recommendations and information abruptly, without accounting for the user’s gradual learning and adaptation process. Learning sciences can help design experiences that create progressive, sustainable paths for shifting attitudes and beliefs toward a process or a tool, rather than applying sudden pressure.

Designing Tools with Regard to the Users’ Behavioral Dimension

As noted earlier, learning involves permanent changes in behavioral potential. Therefore, if a change occurs in the users’ cognition and attitude, we expect to see corresponding changes in their behavioral performance as well.

The learning sciences introduce frameworks to help us understand when and how a behavior becomes consolidated. How, and under what conditions, can we successfully navigate the channel of user cognition, establish a positive attitude toward using a tool or smart test, and ultimately compel them toward stable, purposeful behavior?

The psychology of behavior, as a branch of the learning sciences, steps in at this stage to assist digital tool designers. It prescribes that the formation and continuity of behavioral learning follow specific stages. Therefore, you must clearly define the behavioral prescription you intend to instill in the user:

For instance, in many diabetes management or weight loss programs, merely presenting information about an individual’s health status has not led to behavior change. Studies have shown that when tools provide features for gradual goal-setting, self-monitoring, and continuous feedback, the probability of forming sustainable health behaviors increases.

In short, determine what the target behavior is and define it clearly. Through what stages and micro-steps does this behavior form? What types of feedback and responses guide the user toward the final target behavior? Once the target behavior is formed, what factors or feedback mechanisms can sustain it? And based on what metrics can we determine that the user’s behavior and its continuity result from the proper functioning of the tool?

Furthermore, because our objective is to build human-centered AI and tools, we do not intend to control the user. Instead, we aim to reinforce healthy attitudes and behaviors by boosting their sense of self-efficacy and perceived control over their health journey. In doing so, we guide their behavior and choices along the right path, aligned with the human blueprint.

In fact, one of the growing concerns in the literature on AI in healthcare is the reduction of Human Agency. Some experts have warned that if systems replace human decision-making rather than enhancing it, cognitive dependency and diminished independent judgment may lead to unintended consequences. Hence, the goal of human-centered design must be user empowerment, not user replacement.

Additionally, creating sustainable behavioral habits requires progressive interaction, continuous feedback, and a design that adapts to the real-world context of users’ lives. Tools designed without considering the cognitive and social conditions of the user frequently fail to yield lasting change. Understanding behavioral science and the factors influencing the reinforcement or weakening of a response helps designers correctly guide user behavior while identifying and controlling potential confounding variables inherent in digital tools.

Designing Tools Aligned with the User’s Motivational Dimension

Precisely when some neuroscience specialists argue that everything occurs at the level of cognition and that all other complex human aspects are overshadowed by it, the learning sciences (of which neuroscience is only a part) tell us it is not that simple!!! Learning an idea is a process. If we want a meaningful idea—such as using a health monitoring app—to transform into a highly repetitive, sustained behavior, we must account for other human dimensions as well!

“Trust” is one of the most critical factors in the adoption and motivational continuity of digital health tools. However, trust is not built solely through technical transparency. Users need to feel that the system is understandable, predictable, and psychologically safe.

Several studies have shown that complex or overly technical explanations fail to build trust and instead trigger greater anxiety and confusion among users. Moreover, concerns regarding privacy, data sharing, and the secondary use of health data represent major drivers of distrust among users. In digital health, trust is not merely a technical issue; it is part of the relational experience between the human and the system. For this reason, user experience design must consider psychological safety and relational trust alongside technical security.

Why is trust important? Because it is the loop that connects a user’s cognition, beliefs, and attitudes to their actual behavior! It generates the necessary motivation for follow-through, and ultimately, consolidates a behavior.

Findings from studies conducted on mental health chatbots indicate that anxiety over how personal data is stored, secondary data use, and a lack of transparency regarding data ownership are primary factors driving down user trust. In many instances, users evaluated the perceived quality of the relationship with the system as even more critical than the technical complexity of the algorithm.

While many digital health tools focus heavily on delivering information, possessing information does not automatically translate into the “motivation” required for behavior change. If the user does not feel capable of performing the recommended behavior, the likelihood of continued system utilization drops.

Alongside trust as a motivational component of user behavior, one of the most foundational concepts in the psychology of learning is “self-efficacy.”

Extensive research in health behavior change demonstrates that individuals who believe they possess the capacity to execute recommended actions are far more likely to initiate and maintain the new behavior. Consequently, successful design does not stop at giving advice; it must craft an experience where the user can taste small but meaningful victories.

Self-efficacy refers to an individual’s belief and confidence in their own abilities to organize and execute the courses of action required to achieve a specific goal. This psychological attribute can be modulated via controllable, situational feedback, and AI designers can leverage it as a key lever to impact human motivation.

In a digital health environment, this motivational characteristic can serve as the driving force behind consumer behavior when interacting with smart medical tools. Therefore, AI-driven tools must be capable of reinforcing a sense of empowerment and progressive mastery in the user, rather than merely broadcasting a barrage of alerts and directives. Studies indicate that users achieve more sustainable, satisfying engagement with health systems when feedbacks are personalized, actionable, and contextualized within the actual reality of their daily lives.

Feedback itself is only effective when it is timely, clear, and meaningful. Generic, non-actionable feedback—such as vague lifestyle advice—typically exerts a highly limited impact on behavior change. Conversely, contextualized, action-oriented feedback can significantly heighten both the cognitive and motivational engagement of the user.

Expected Operational Implications for Design and Development Teams

Many digital health tools still focus predominantly on algorithmic performance and technical functionalities, whereas the sustainability of human-system interaction relies on the quality of the cognitive and behavioral experience—and, of course, the motivational loop that links cognition to behavior and drives its continuity.

The learning sciences and psychology of learning can empower design and development teams to move far beyond the mere metrics of “ease of use” and “time management.” This shift is the most vital achievement of a cognitive, motivational, and behavioral architecture governing human-AI interaction.

This issue holds particular urgency for startups, product design teams, and digital health developers; the true success of these tools does not hinge on the sheer number of features, but on their capacity to preserve sustained engagement, secure trust, and guide human behavior change.

The learning sciences offer a framework to design tools that are not just usable, but comprehensible and justifiable within the users’ cognitive schemas. It shapes and directs user behavior under their own autonomy through self-regulation mechanisms, supplies the motivational loops connecting thought to action, and stabilizes user behavior.

Perhaps the most definitive question for the future of digital health is not how much smarter artificial intelligence will become, but rather how much better it can comprehend human cognition, motivation, agency, and behavior. Ultimately, the most successful tools will not necessarily feature the most advanced algorithms, but those that possess the deepest understanding of the human being.

References

Bandura, A. (1997). Self-Efficacy: The Exercise of Control. New York: W.H. Freeman.

Schunk, D. H. (2020). Learning Theories: An Educational Perspective (8th ed.). Pearson.

Hergenhahn, B. R., & Olson, M. H. (2015). Theories of Learning (7th ed.). Pearson.

Zimmerman, B. J. (2002). “Becoming a Self-Regulated Learner: An Overview.” Theory Into Practice, 41(2), 64-70.

Deloitte. (2024). 2024 Global Health Care Outlook.

Deloitte Center for Health Solutions. (2024). Digital Transformation and Consumer Engagement in Healthcare.

Rock Health. (2024). Digital Health Consumer Adoption Survey.

Blease, C., Kaptchuk, T. J., Bernstein, M. H., et al. (2019). “Artificial Intelligence and the Future of Primary Care.” The Lancet Digital Health, 1(8), e353-e354.

Topol, E. (2019). Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again. Basic Books.

World Health Organization (WHO). (2021). Ethics and Governance of Artificial Intelligence for Health.

Sweller, J. (1988). “Cognitive Load During Problem Solving: Effects on Learning.” Cognitive Science, 12(2), 257-285.

The post Operationalizing Learning Sciences for Human-Centered AI in Digital Health appeared first on Medika Life.

At a time when Artificial Intelligence is rapidly weaving itself into the fabric of physical and mental healthcare, basic user-friendliness, processing speed, and market acceleration are no longer enough. To build digital solutions that actually work, we must grasp how humans learn, adapt, and transform their behaviors. This is exactly where the learning sciences become vital. Put simply, until we decode the mechanisms of ‘deep learning in humans’ through the lens of learning sciences, the concept of ‘Deep Learning’ in AI development will never reach its true potential.

Digital health, much like any other modern domain, is now permanently tied to technology. From education and corporate structures to parenting and economics, technology is built to streamline processes, widen access, and boost precision. At its core, technology was created to serve humanity across individual and social spheres, and digital health stands as one of the most critical testing grounds for this promise.

Yet, alongside this reality lies a much bigger issue—one that is gaining traction and deserves a rigorous, interdisciplinary look.

The question isn’t whether technology is inherently good or bad; it is that even the most advanced technology remains ineffective if it fails to align with human blueprints.

Today, more than ever, we need to look at AI and digital systems through a deeply human lens. This means moving away from treating an individual merely as a ‘user,’ a ‘data processor,’ or a passive ‘receiver,’ and instead recognizing them as a multi-dimensional, complex, living being.

In digital health, our core focus is the human being—the patient striving for recovery, the client seeking a precise diagnosis, the therapist requiring sharper diagnostic tools, or the physician leaning on technology to make high-stakes clinical decisions. The human is always the ultimate destination. If a digital tool is to succeed in this space, it must genuinely connect with real people, accounting for their cognitive, behavioral, biological, and experiential complexities.

Why Research in the Learning Sciences is Indispensable

In the digital health space, the real challenge is never just about getting someone to install an app or use a digital tool temporarily. The true measure of success is whether that tool can drive a real, lasting change in human behavior, attitude, and lifestyle. If a person engages with a platform for a brief period but experiences no sustainable shift in their health or daily habits, the technology has fundamentally missed its mark.

This is where the learning sciences help us elevate technology design far beyond surface-level mechanics and computational algorithms. When we understand how a person actually internalizes information, we can build better communication strategies, deliver more constructive feedback, apply the right behavioral reinforcements, and create environments that foster genuine trust, motivation, and user engagement.

Furthermore, this scientific backing allows us to grasp privacy and data security from the psychological standpoint of the user, since a patient’s willingness to trust a system is directly tied to how safe they feel sharing their data.

Two Foundational Pillars: Trust and Continuance Intention

To see how the learning sciences practically guide human behavior in the era of AI, we can look at two crucial dynamics in digital health:

1. The Mechanics of Trust
Trust is the ultimate currency in digital health, because users are asked to hand over highly sensitive personal, biological, and psychological data to an algorithm.

2. Continuance Intention and Habit Formation
Capturing a user’s attention at launch is relatively easy; keeping them engaged over time is where the tech industry routinely struggles.

Conclusion

The defining critique of modern AI is not its widespread adoption, but its lack of authentic human-centricity. Successful digitalization in healthcare cannot rely solely on technical scalability; it must place the complex human being squarely at the center of the design process.

Technology only gains meaning when it can understand human beings, build a relationship with them, earn their trust, and guide them toward lasting well-being.

Ultimately, the future of digital health will not be measured by raw processing power, but by the depth of the developer’s understanding of the human condition.

References

Sucala, M., Cole-Lewis, H., Arigo, D., Oser, M., Goldstein, S., Hekler, E. B., & Diefenbach, M. A. (2021). Behavior science in the evolving world of digital health: Considerations on anticipated opportunities and challenges. Translational Behavioral Medicine, 11(2), 495–503. https://doi.org/10.1093/tbm/ibaa034

Bai, B., & Guo, Z. (2022). Understanding users’ continuance usage behavior towards digital health information system driven by the digital revolution under COVID-19 context: An extended UTAUT model. Psychology Research and Behavior Management, 15, 2831–2842. https://doi.org/10.2147/PRBM.S364275

The post Human-Centered AI in Digital Health: Why Learning Sciences Matter appeared first on Medika Life.

This white paper was researched and written with the assistance of Claude Sonnet, an AI system developed by Anthropic. AI assistance was used to accelerate literature retrieval, improve the quality of writing, and support editing and formatting. The intellectual framework, argument structure, source selection, and all substantive claims reflect the author’s own thinking and direction. All citations have been identified and verified by the author. The author assumes full responsibility for the accuracy and integrity of all content presented in this paper.

Executive Summary

Artificial intelligence is arriving in American healthcare at scale. Health systems are investing in AI-powered diagnostics, clinical decision support, predictive analytics, and administrative automation. The promise is real. So is the risk. Machine learning models learn from data. In healthcare, that data is generated by the systems deploying the AI. And if those organizations have not been designed to produce clean, reliable, clinically meaningful data, then the AI built on top of them will automate and amplify the dysfunction already present in the system, not correct it.

This is the argument this paper makes. It is not primarily an argument about technology. It is an argument about organizational design.

The concept of the Learning Health System, formally defined by the Institute of Medicine in 2007, describes a system in which knowledge generation is so deeply embedded in the delivery of care that improvement becomes continuous and self-reinforcing rather than episodic and externally driven. Nearly two decades after that definition was published, widespread adoption remains limited. The gap is not one of awareness. It is one of operationalization. And in an era of AI-driven healthcare, the cost of that gap is no longer just missed improvement opportunities. It is corrupted training data, biased models, and clinical decisions shaped by intelligence that learned the wrong things from a system that was never designed to learn at all.

This paper examines why the Learning Health System has not been built at scale, using the organizational thinking design framework of Vision, Mission, Capacity, and Learning developed by Drs. Derek and Laura Cabrera, and the wicked problem literature in strategic management. It identifies three conditions most visible in clinical, policy, and public discourse as illustrations of the organizational design problem: physician burnout, electronic health record burden, and payer interference through prior authorization. These three are not presented as an exhaustive explanation. They are presented as a coherent causal chain that leads directly to the data quality crisis sitting underneath every AI deployment in American healthcare today.

The paper concludes not with a prescriptive framework but with an invitation to think differently about how health systems are designed, led, and held accountable, before the next wave of AI investment locks in the mistakes of the current one.

I: A Conversation That Sparked a Question

American healthcare is in the middle of an AI gold rush. Health systems, technology companies, and investors are moving fast, betting that machine learning, predictive analytics, and AI-powered clinical tools will transform how care is delivered and how outcomes are measured. The enthusiasm is understandable. The technology is genuinely powerful. But a question is not being asked loudly enough: what kind of system is this AI learning from?

In early 2026, Gil Bashe, Chair of Global Health and Purpose at FINN Partners, published Healing the Sick Care System: Why People Matter, arguing that American healthcare is not failing because it lacks innovation, investment, or talented people.[2] It is failing because it has lost sight of the people it exists to serve. That argument sparked a different but related question for the author: what kind of system do we actually have?

We call them healthcare systems. We build teaching hospitals. We invest in teaching rounds and residency programs and the careful, structured transmission of clinical knowledge from one generation to the next. Teaching is a word we use with confidence and pride in medicine. But when do we talk about the system itself learning? Not individuals acquiring competency, but the institution changing what it does based on what it discovers. Teaching and learning are not the same thing, and that distinction, hiding in plain sight, may be one of the most consequential unexplored ideas in American healthcare today, especially at a moment when AI is being asked to learn from systems that were never designed to learn themselves.

This question led to an examination of a concept that has existed in formal academic and policy literature since 2007 but has not entered the broader conversation about healthcare reform in any meaningful way: the Learning Health System.

II: What Is a Learning Health System, and Why Has It Not Been Built?

Understanding why AI in healthcare is sitting on a compromised foundation requires understanding what a Learning Health System actually is, and why one has never been fully built. The Learning Health System is not simply a framework for improving data quality. It is the only organizational model in which clean, clinically meaningful data is a natural and continuous byproduct of how care is delivered. Every other approach to the data quality problem in healthcare AI is essentially trying to fix the output without changing the system that produces it. The Learning Health System changes the system. That is why it matters now, and that is why AI in healthcare makes it urgent.

The term Learning Health System entered the formal vocabulary of American medicine in 2007 when the Institute of Medicine convened a roundtable on value and science-driven health care. The definition it produced has held up well: a Learning Health System is one in which knowledge generation is so embedded into the core of the practice of medicine that it is a natural outgrowth and product of the healthcare delivery process and leads to continual improvement in care.[1] Knowledge generation in this vision is not adjacent to practice. It is not a research department down the hall or a quality improvement initiative launched when funding permits. It is embedded in practice itself, and it leads to continual, self-reinforcing improvement in which care creates evidence and evidence improves care.

Nearly two decades later, widespread adoption remains limited. Not because the concept has been ignored. It has attracted sustained attention from the National Academy of Medicine, federal agencies including Agency for Healthcare Research and Quality (AHRQ) and Patient-Centered Outcomes Research Institute (PCORI), major academic health centers, and research networks such as National Patient-Centered Clinical Research Network (PCORnet) and the NIH’s National COVID Cohort Collaborative. What has proven difficult is operationalization at scale: figuring out what a genuine commitment to learning actually means in terms of changed practice, realigned infrastructure, new staffing, revised policy, and real shifts in organizational culture. The IOM’s deliberately broad definition, intended to maximize applicability, had an unintended consequence. It left every institution to solve the operationalization problem largely on its own, without a shared language for the organizational design work that learning at scale actually requires.[16]

The cycle the Learning Health System literature describes is straightforward in concept. Knowledge is identified and synthesized to address clinical challenges through evidence reviews and clinical practice guidelines. That knowledge gets applied in care delivery through clinical decision support and care pathways. Care delivery generates data, captured in patient registries and EHRs, assessed for performance, and fed back into the knowledge generation process. The loop closes. Patients are at the center throughout, not as passive recipients of decisions made elsewhere, but as active contributors to the knowledge the system generates.[11]

It is also worth being clear about what a Learning Health System is not. It is not a teaching hospital. A teaching hospital organizes itself to transfer knowledge from experienced clinicians to trainees. Knowledge flows in one direction, and the institution learns incidentally if at all. A Learning Health System organizes itself to change based on what it discovers in the course of delivering care. The institution itself is the learner. American medicine has invested heavily in building teaching capacity. The investment in learning capacity, the organizational infrastructure that allows a health system to discover, synthesize, and act on what its own practice is telling it, has been far more limited and far less systematic.

The concept operates at two levels that are easy to conflate. At the macro level, it describes what American healthcare as a sector could become. At the micro level, it is an organizational design challenge that has to be solved institution by institution through specific decisions about how care is delivered, how data is captured, how knowledge is synthesized, and how evidence actually changes what clinicians do on any given day. The macro vision only becomes real through micro organizational choices. The research literature suggests those choices have not yet been made in ways that support learning at meaningful scale.

III: A Wicked Problem and a Strategic Dilemma

Before examining why the Learning Health System has been so difficult to build, it is worth being precise about the nature of the problem itself. Not all hard problems are the same kind of hard. Some are difficult because resources are insufficient. Some are difficult because the right solution has not yet been found. The failure to operationalize the Learning Health System at scale is neither of these. It is something more structurally challenging, and naming it correctly matters because the type of problem determines what kind of thinking is adequate to address it.

In strategic management and organizational theory, a distinction is drawn between problems that are complicated and problems that are wicked. A complicated problem, however technically demanding, has a definable solution. Building an aircraft is complicated. The right answer exists, the variables can be enumerated, and expertise applied systematically will eventually produce the result. A wicked problem is different in kind, not just in degree. The concept was introduced by Rittel and Webber in their foundational 1973 paper “Dilemmas in a General Theory of Planning,”[5] which argued that problems of social policy cannot be solved using scientific-engineering approaches because they lack a clear problem definition and involve stakeholders with genuinely differing and legitimate perspectives. Wicked problems are not merely unsolved. They resist definitive formulation. Every attempt to solve them reveals new dimensions of the problem. Solutions cannot be tested in advance and cannot be undone cleanly once implemented. There is no single right answer, and the people working on the problem do not agree on what success would look like.

The challenge of building a Learning Health System is a wicked problem in precisely this sense. It is not a technology problem, though technology is implicated. It is not a regulatory problem, though regulation shapes the environment. It is not a funding problem, though funding matters. It is a problem that cuts across all of these domains simultaneously, involves stakeholders whose legitimate interests are in genuine tension with one another, and resists any solution that addresses only one of its dimensions. Researchers working in this space have noted that strategy scholars who attempt to address wicked problems using conventional approaches tend to build causal models that seek to optimize organizational success, an approach that ironically divorces the analysis from the very complexity that makes the problem wicked in the first place.[6]

Within this wicked problem, however, there is a more specific structure worth naming. The Learning Health System presents what might be called a strategic dilemma: a situation in which legitimate goods are in genuine tension with each other, and in which choosing to prioritize one value necessarily creates pressure on another. Patient safety and the imperatives of research require different things from a consent framework. The need for standardization conflicts with the need for clinical judgment. The value of data utility for population-level learning conflicts with individual privacy rights. The urgency of improvement conflicts with the rigor that improvement based on evidence requires. These are not tensions that can be dissolved by finding a smarter solution. They are structural features of the problem that any serious approach must hold in view simultaneously rather than resolving prematurely in favor of one side.

This distinction between a wicked problem and a strategic dilemma is not merely academic. It has direct implications for how we think about leadership and organizational design in this space. Wicked problems cannot be assigned to a committee and solved on a timeline. They require what the Cabreras would describe as thinking design rather than framework imposition: the cultivation of a quality of thinking in leaders and institutions that is capable of holding complexity, adapting continuously, and learning from the system rather than simply managing it. The Learning Health System is not waiting for the right policy. It is waiting for a different quality of organizational thinking. And that is a problem that systems thinking, properly understood, is specifically designed to address.

IV: Organizations as Complex Adaptive Systems — The Cabrera Lens

Understanding why the Learning Health System has been so difficult to operationalize requires more than a catalogue of obstacles. It requires a way of thinking about organizations that is adequate to their actual nature. Most health systems have been designed and managed as if they were complicated machines: hierarchical, controllable, and optimizable through the right combination of process improvement, technology, and incentive alignment. The persistent failure of that approach to produce genuine organizational learning suggests that the underlying model of what a health system is may itself be the problem.

Drs. Derek and Laura Cabrera at Cabrera Research Lab have spent decades developing and empirically grounding a different model. Their work, elaborated in Flock Not Clock and in an extensive body of peer-reviewed research,[3] begins from a foundational premise: all organizations, regardless of their formal structure, are complex adaptive systems. A complex adaptive system, or CAS, is composed of autonomous agents whose individual behaviors interact to produce collective, emergent outcomes that cannot be predicted or controlled by managing the agents individually.[13] The agents are not cogs in a machine executing instructions from above. They are people making decisions, moment by moment, in response to the conditions and incentives around them. The organization does not produce its outcomes by command. It produces them by emergence, as the aggregate result of countless individual decisions made at every level of the system every day.

This changes how we think about organizational design. If a health system is a complex adaptive system, then the question of how to build a learning culture inside it is not primarily a question of policy, technology, or incentive structure, though all of these matter at the capacity level. It is a question of what conditions and orientations the autonomous agents in the system are operating under, and whether those conditions make learning a natural emergent outcome of their daily work or an additional burden layered on top of everything else they are already asked to do.

The Cabreras developed a thinking design structure called VMCL, standing for Vision, Mission, Capacity, and Learning, to help leaders understand and shape the four functions that any organization must perform in order to move purposefully toward its goals.[4] VMCL is not a framework to be implemented as a checklist or adopted as a rebranding exercise. It is a thinking design lens, a way of seeing clearly what an organization is actually doing across its four essential functions, and whether those functions are genuinely aligned with each other and with the organization’s deepest purpose. The value is in the quality of thinking it cultivates in leaders, not in the mechanical application of its categories. Of the organizational design frameworks the author has encountered across three decades of operational leadership, the Cabrera VMCL structure is the most useful for making visible what is actually happening inside a complex organization and why.

Vision is a destination, not an action. It is a picture of a specific future state, clear enough to be genuinely directional and distant enough to be genuinely aspirational. Vision is not a description of what the organization does or how it operates. It is the answer to the question: if everything this organization is trying to accomplish were fully realized, what would the world look like? Most organizational vision statements fail this test entirely. They are the product of committee processes in which boards, executives, communications professionals, and legal reviewers each add words until the original impulse toward meaning has been buried under qualifications and compromises. The result is statements that are long, passive, and forgettable, that could belong to any organization and therefore belong to none, and that no frontline worker could honestly say lives in their hearts and minds while doing their job. Genuine vision is short enough to remember, true enough to feel, and clear enough to orient behavior without requiring a footnote.

Mission is the mechanism by which vision becomes real. In the VMCL structure, mission is not a values statement or a description of organizational purpose. Mission is the simple rules: the small number of repeatable, measurable actions that, when enacted consistently by autonomous agents throughout the organization, produce movement toward the vision as an emergent outcome.[12] The Cabreras draw on complex adaptive systems science to make a counterintuitive but empirically grounded argument: large-scale coordinated behavior in complex systems does not require elaborate instructions or top-down control. It requires simple rules, followed by many agents, repeatedly. Consider the wave at a stadium. No policy memo was issued. No training was conducted. The behavior that ripples across tens of thousands of people in a single coordinated arc emerges from a small number of simple rules enacted by each individual: watch your neighbor, rise when they rise, sit when they sit, raise your hands. The wave is not managed into existence. It emerges. Mission, properly conceived, functions the same way inside organizations. If the simple rules of mission are well designed, genuinely understood, and authentically shared, coordinated movement toward vision emerges from the collective behavior of autonomous agents without requiring command and control of every decision. The parallel failure mode matters equally: if mission consists of a lengthy statement written for external audiences rather than a small number of actionable rules that people can actually carry in their heads, then the organization’s agents have nothing simple to enact, and the coordinated movement that vision requires cannot emerge.

Capacity is the infrastructure, systems, tools, skills, and resources that enable the mission to be carried out. It is what the organization has built, or inherited, or been forced to adopt, to allow its agents to do the work that produces the vision. Capacity includes technology, physical infrastructure, trained personnel, financial resources, data systems, and organizational structures. The critical insight in the VMCL framework is that capacity must be aligned with mission. Capacity built for a different mission, however large, sophisticated, or expensive, does not support the mission it was not designed to serve. It actively competes with it, consuming the time, attention, and energy of the autonomous agents who are supposed to be carrying out the simple rules that produce the vision. The question of whether a health system has the capacity to be a Learning Health System is therefore not simply a question of whether it has electronic health records, data analytics capabilities, or quality improvement staff. It is a question of whether those investments were designed and are being used in service of a learning mission, or whether they were designed for other purposes entirely and are now being asked to serve a mission they were never built to support.

Learning is the function that makes the other three adaptive rather than static. In the VMCL framework, learning is the organization’s capacity to gather honest feedback from its own behavior and from its environment, assess that feedback against its vision and mission, and actually change what it is doing as a result.[4] In the specific context of the Learning Health System, this has a precise meaning that goes beyond general organizational learning or individual professional development. Learning in the LHS sense is the cycle of gathering clinical and operational data generated within the health system itself, subjecting it to rigorous analysis, producing knowledge about what is actually working for actual patients in this actual system, and feeding that knowledge back into changed clinical practice in ways that improve patient outcomes. The unit of learning is the system. The measure of learning is not the number of insights generated or reports published. It is whether practice changes and whether patients do better as a result. Quality dashboards that nobody acts on, annual reports that circulate among administrators without altering clinical behavior, and research findings that never make it from the journal to the bedside are all symptoms of an organization that has the appearance of learning without the substance of it.

These four functions are not sequential steps. They are simultaneous and mutually dependent. Vision without mission produces inspiring rhetoric that changes nothing. Mission without vision produces activity without direction. Capacity without aligned mission and vision produces expensive infrastructure that serves the wrong ends. And Learning without the other three produces insight that has no home in the organization’s structure and no pathway to changing behavior. The question the VMCL lens asks of any health system is not whether these four functions exist in some form, because they all do in every organization. The question is whether they are genuinely aligned with each other, whether they are all oriented toward the same destination, and whether that destination is honestly about learning and patient outcomes or about something else dressed in that language.

V: Three Conditions Hostile to Learning

The VMCL lens developed by the Cabreras does not merely describe what a well-functioning organization looks like. It also provides a diagnostic structure for understanding where and why organizational function breaks down. When a complex adaptive system is failing to move toward its vision, the failure can almost always be located in one or more of the four functions: the vision is unclear or not genuinely shared, the mission lacks simple rules that agents can actually carry and enact, the capacity is misaligned with the mission, or the learning function is absent, performative, or structurally disconnected from the decisions that govern practice.

Applied to the challenge of building Learning Health Systems in the United States, this diagnostic structure surfaces something important. The barriers most frequently discussed in clinical, policy, and public discourse cluster with particular intensity around the Capacity and Learning functions. Three conditions in particular have emerged with enough consistency across enough professional, policy, and clinical circles to warrant focused examination here. They are not presented as the only barriers. The published literature names others, including interoperability failures, governance gaps, funding misalignment, and cultural resistance to change.[15] They are presented because each is vivid, well-documented, and together they do something more important than illustrate three separate problems. They form a causal chain.

That chain runs as follows. Electronic health record systems were designed for billing, documentation, and regulatory compliance rather than for clinical care or learning. They impose structural friction on the daily work of every physician in the country. Payer interference through prior authorization requirements compounds that friction, consuming hours of clinical time every week, systematically overriding clinical judgment, and producing a persistent experience of professional constraint that no amount of individual resilience can fully absorb. Together these two systemic forces create the organizational conditions that produce physician burnout at scale. Burnout is not an independent variable sitting alongside EHR burden and payer interference. It is the human output of a system that has been designed at the capacity level for the wrong mission. And a system whose agents are burned out cannot learn, because learning requires the cognitive availability, the reflective capacity, and the institutional trust that survival mode structurally forecloses.

This is what the Cabreras mean when they say that the system is what the system does. If the system consistently produces burned-out physicians, demoralized care teams, and a clinical workforce increasingly oriented toward self-preservation rather than adaptive engagement, that is not a failure of individual character or professional commitment. It is the system performing as it was designed to perform, optimizing for throughput, administrative control, and reimbursement rather than for learning and patient outcomes. Understanding the three conditions in sequence, rather than as a parallel list, is essential to understanding why the organizational design problem is as deep as it is.

Electronic Health Records: Capacity Built for the Wrong Mission, Sitting on the Right Data

The widespread adoption of electronic health records in the United States was accelerated by the Health Information Technology for Economic and Clinical Health Act of 2009 [23]. As of 2021, 96 percent of nonfederal acute-care hospitals and 78 percent of office-based physicians used an EHR, making these systems integral to routine clinical practice.[10] On its face, this represents exactly the kind of data infrastructure that a Learning Health System requires. A system that captures clinical data at scale, across encounters, patients, and populations, is precisely what the knowledge generation and data functions of the LHS cycle depend on. In this narrow sense, American healthcare has already built something the Learning Health System needs. The data is there. Decades of patient encounters, clinical decisions, treatment courses, and outcomes are sitting in these systems at a scale that would have been unimaginable to the architects of the NAM’s 2007 vision.

The problem is not the existence of the data. The problem is everything surrounding it.

EHRs were not primarily designed for learning. They were designed for billing, documentation, and regulatory compliance. The gap between the data infrastructure a learning mission requires and the data infrastructure that exists is not a gap in hardware or software capability. It is a gap in design intent, and that gap has consequences that run in two directions simultaneously. The first is the burden the systems impose on the clinicians who must feed them. A recent scoping review published in the Journal of Evaluation in Clinical Practice found that clinicians now spend an estimated one-third to one-half of their working day interacting with EHR systems, translating to over $140 billion in lost care capacity annually.[10] The same review found that clinicians frequently experience significant workflow disruptions caused by poorly designed interfaces, leading to task-switching, excessive screen navigation, and fragmented critical information that necessitates workarounds and increases the risk of documentation errors. Research published in JAMA found that physicians spend approximately 36.2 minutes documenting in the EHR for every 30-minute office visit [24], meaning the administrative burden of capturing an encounter now routinely exceeds the clinical time of the encounter itself.

The second consequence is less frequently discussed but equally important for the Learning Health System argument. The data that EHRs generate is not clean learning data. It is documentation data, structured around billing codes, shaped by prior authorization requirements, and produced through documentation processes that clinicians have adapted, often through workarounds, to minimize burden rather than to maximize clinical accuracy. The result is a paradox at the heart of the LHS challenge: American healthcare is sitting on an extraordinary volume of clinical data that a learning system would need, and simultaneously that data is less useful for learning than its volume suggests, because the processes that generated it were optimized for reimbursement rather than for clinical fidelity.

Mining that data for genuine learning insights would require significant investment in data science, informatics, and clinical expertise working in close collaboration. It would require clinicians who have the time, the cognitive availability, and the institutional support to participate in that work. It would require organizations that have aligned their capacity with a learning mission rather than a billing mission. And it would require a workforce that has not been burned out by the very systems that are generating the data in the first place. The EHR is not an obstacle to the Learning Health System in spite of the data it holds. It is an obstacle in part because of the conditions it has created around that data. The data exists. The capacity to act on it does not, because the system has consumed that capacity in the process of generating the data.

In VMCL terms this is a Capacity problem of a specific and frustrating kind. The investment has been made. The infrastructure is in place. But it was built for the wrong mission, and the friction it generates spills directly into the clinical encounter itself, into the relationship between physician and patient, and into the professional experience of every clinician who ends the day staring at a screen long after the last patient has gone home.

Payer Interference: External Rules Overriding Internal Mission

If EHR burden creates structural friction in the tools physicians use, payer interference through prior authorization creates structural friction in the decisions physicians are permitted to make. Together they constitute a double compression of clinical capacity that is difficult to fully appreciate from outside the daily experience of practicing medicine in the United States today.

The American Medical Association conducts an annual nationwide survey of 1,000 practicing physicians on the burden of prior authorization. The 2024 findings are both consistent with prior years and striking in their severity.[9] Physicians reported completing an average of 39 prior authorization requests per physician per week, consuming an average of 13 hours of physician and staff time. Ninety-three percent of physicians reported that prior authorization delays access to necessary care. Eighty-nine percent reported that it contributes to burnout. Ninety-four percent said it has a negative impact on patient clinical outcomes. More than one in four reported that prior authorization caused a serious adverse event for a patient in their care. Seventy-eight percent reported that it often or sometimes results in patients abandoning a recommended course of treatment entirely. Forty percent of practices have hired staff whose exclusive function is managing prior authorization requests.

In the language of complex adaptive systems, prior authorization represents external agents, payers and insurers, injecting rules into the system that redirect the behavior of internal agents, physicians and care teams, away from what their clinical training, judgment, and the available evidence would support, and toward what the external agent will reimburse. The internal simple rules of the care delivery mission are being overridden at the point of care by administrative requirements that serve a different set of goals entirely. This is not a marginal disruption. At 39 prior authorization requests per physician per week, it is a structural feature of the environment in which clinical work now happens.

The implications for the Learning Health System extend beyond the administrative burden. The LHS cycle depends on clinical practice generating data that reflects actual clinical judgment applied to actual patient needs. When a substantial proportion of clinical decisions are being shaped not by evidence and judgment but by prior authorization requirements, the data that clinical practice generates no longer cleanly reflects what works. It reflects what gets approved. The knowledge that a learning system could generate from that data is therefore systematically biased before it is ever analyzed. The learning loop is not merely slowed by payer interference. In important respects it is compromised at the source.

And when a physician has spent 13 hours in a week on prior authorization paperwork, on top of the hours already consumed by EHR documentation, the cumulative weight of that friction does not remain a professional inconvenience. It becomes a clinical emergency of a different kind entirely. It becomes burnout.

Physician Burnout: The Human Output of a Broken System

Physician burnout is not the beginning of the problem. It is the end of a chain that starts with organizational design decisions made far from the bedside. It is what happens when the agents of a complex adaptive system are placed inside a capacity structure so misaligned with the mission of care that adaptive engagement becomes unsustainable. The EHR consumes time and cognitive energy. Prior authorization consumes professional agency and clinical judgment. Together they produce a working environment in which the question a physician must increasingly ask is not what does this patient need but what will I be permitted to do, and how long will the paperwork take.

The data on physician burnout in the United States is not ambiguous. According to the Dr. Lorna Breen Heroes’ Foundation, 76 percent of healthcare workers reported burnout in 2020, and during the COVID-19 pandemic 69 percent of physicians experienced depression, with 13 percent reporting thoughts of suicide.[7] Physicians in the United States are more likely to die by suicide than physicians in other nations. The Physicians Foundation’s 2022 Survey of America’s Physicians found that burnout rates remain at 62 percent, significantly higher than the pre-pandemic figure of 40 percent in 2018, with no meaningful improvement in the intervening years.[8] Nearly 400 physicians die by suicide annually in the United States, a figure the research literature connects directly to stigma, fear of licensing repercussions, and untreated depression in a profession that has historically treated the need for mental health support as a professional liability.[7]

The Dr. Lorna Breen Heroes’ Foundation, established by the family of an emergency physician who died by suicide in April 2020 after treating patients during the early COVID-19 surge, has been explicit about the systemic nature of the problem. Individual support alone, the foundation states, does not address the causes of burnout. The underlying processes and systems within healthcare operations must be confronted.[7] That is a systems thinking argument made in plain language by people who lived the consequences. It points directly at the Capacity layer of the VMCL structure and asks why the system was designed this way and whether the people responsible for that design have fully reckoned with what it produces.

For the Learning Health System, burnout represents the final compression of capacity. Learning requires clinicians who can observe, reflect, contribute to knowledge generation, and adapt their practice in response to what the evidence is telling them. It requires agents who are present, engaged, and operating with enough cognitive and professional reserve to participate in something beyond the immediate transaction of care. Burnout forecloses that participation systematically, across specialties, settings, and the full arc of a clinical career. A system that is burning out its physicians at the rate American healthcare currently does is not a system that can learn. It is a system that is consuming its own capacity to improve.

The three conditions examined in this section are not a complete explanation of why Learning Health Systems have been so difficult to build. But they are a coherent one. They describe a system that has built the wrong capacity, allowed that capacity to be further distorted by external rule-making, and in doing so created the organizational conditions that make the human beings at the center of care less and less able to participate in the continuous learning that better care requires. The system is, in the most precise sense, doing exactly what it was designed to do. The question this paper is asking is whether it could be designed to do something different.

VI: Thinking Design, Not Framework Prescription

If the argument of this paper has been constructed carefully, the reader has arrived here with a specific kind of discomfort. The problem is real, well-documented, and serious. The VMCL lens has provided a coherent way of seeing why the Learning Health System has not been built at scale. The three conditions examined in Section V have illustrated, in concrete and citable terms, how the capacity layer of American healthcare has been so comprehensively misaligned with a learning mission that the human beings at the center of care are being systematically consumed by the friction of a system that was designed for other ends. The natural next question is: so what do we do about it?

This section is going to resist the impulse to answer that question with a prescription. That resistance is not evasion. It is the most honest and useful response available, and the reasons for it are worth stating plainly.

The wicked problem literature is clear that conventional problem-solving approaches are structurally inadequate to problems of the kind this paper has been examining. The Learning Health System is not waiting for the right policy intervention or the right technology platform or the right reimbursement model, though all of these matter and deserve serious attention. It is waiting for a different quality of organizational thinking in the people and institutions responsible for designing, leading, and reforming American healthcare.

The Cabreras make a distinction that is useful here. They differentiate between organizations that impose frameworks and organizations that develop genuine thinking capacity, the internal ability to see clearly, reason carefully, and adapt continuously in response to what the system is actually doing.[3] A framework can be adopted without changing the underlying quality of thought. A new software platform can be installed without changing the organizational culture that will use it. A new policy can be passed without changing the incentive structures that will determine whether it is followed in spirit or circumvented in practice. What cannot be faked, and what the Learning Health System actually requires, is the organizational capacity to ask honest questions about what the system is producing, to follow the answers wherever they lead, and to change course based on what is discovered.

Before any of that can happen, the system must be mapped. Not fixed. Not optimized. Mapped. This is a critical distinction. The problems do not precede the mapping. They emerge from it. A system cannot be improved by agents who cannot see it clearly, and seeing it clearly requires a specific and disciplined quality of thinking. The Cabreras offer exactly that through a cognitive framework called DSRP, standing for Distinctions, Systems, Relationships, and Perspectives.[19][21] DSRP describes four universal patterns of thinking that, when applied deliberately, allow a leader or organization to see a system as it actually is rather than as habit, assumption, or organizational mythology would have it appear. To understand what the system does, you must first understand what the system is. DSRP is the toolkit for that work.

Before reaching for solutions, the Cabreras ask leaders at every level to sit with a set of honest diagnostic questions:

Does your organization have a vision that is genuinely and specifically about the future it is trying to create, stated clearly enough that every person in the system, from the bedside nurse to the chief executive, could carry it in their hearts and minds while doing their job on any given day? Or does it have a statement written for a board presentation, long, passive, and laden with qualifications, that could belong to any organization and therefore belongs to none?

Does your organization have a mission in the specific sense of simple rules, repeatable actions that autonomous agents at every level of the system can enact without a manual, that would make learning a natural outgrowth of daily clinical practice? Or does it have a strategic plan, full of initiatives and objectives and key results, that bears no relationship to what a nurse or a physician or a data analyst actually does on a Tuesday morning?

Has your organization built capacity that is aligned with a learning mission, or has it built capacity for billing, documentation, and regulatory compliance and then asked that infrastructure to support learning as a secondary function while simultaneously burning out the people who are supposed to use it?

And does your organization have genuine learning mechanisms, honest feedback that actually changes clinical practice, that actually improves patient outcomes, that actually closes the loop between what the system discovers and what the system does? Or does it have quality dashboards and compliance reports and annual reviews that circulate among administrators without ever altering what happens in an exam room?

These are diagnostic questions, not rhetorical ones. They are the questions that thinking design asks of any organization that claims the Learning Health System as an aspiration. They are uncomfortable because for most health systems, across most of these dimensions, the honest answer is not encouraging. And they are important precisely because the discomfort they produce, if it is held rather than resolved prematurely, is the beginning of genuine organizational learning.

The four DSRP patterns work as follows.

Distinctions are the act of identifying what something is and what it is not, drawing a boundary between a thing and everything that is not that thing. In the context of the Learning Health System, making clear distinctions means being honest about what a learning system actually is, and separating it clearly from what merely resembles it. A teaching hospital is not a learning health system. A quality dashboard is not a learning mechanism. An EHR is not a learning infrastructure simply because it generates data. Without the discipline of making clean distinctions, organizations substitute the appearance of learning for the substance of it and never notice the difference.

Systems, in the DSRP sense, is the recognition that any phenomenon of interest is simultaneously a part of larger wholes and a whole composed of smaller parts, and that understanding it requires attending to both levels at once.[20] In the healthcare context, physician burnout is a part of a larger system of capacity failures, and it is itself a whole composed of contributing conditions including EHR burden, prior authorization load, professional isolation, and the erosion of clinical agency. Understanding both the part and the whole simultaneously is what prevents the mistake of treating burnout as an individual problem rather than a systemic one.

Relationships are the causal and dynamic connections between elements of a system, the action and reaction that link one condition to another and produce the emergent outcomes the system generates.[20] The causal chain this paper has traced, from EHR misdesign through payer interference to burnout to the collapse of learning capacity, is a relationships argument. These three conditions are not parallel and independent. They are sequentially and causally connected, and intervening in one without attending to the others will produce incomplete and temporary relief at best.

Perspectives are the recognition that every observation of a system is made from a point of view, and that changing the perspective from which a system is examined reveals different features, different problems, and different possibilities.[20] The Learning Health System has been examined primarily from the perspectives of bioethicists, health policy scholars, and informatics researchers. Those are valuable perspectives. But they are not the perspective of the burned-out emergency physician at the end of a 13-hour shift, or the primary care doctor who spent two of those hours on prior authorization paperwork, or the patient whose recommended treatment was abandoned because the approval process took too long. Bringing multiple genuine perspectives into the analysis is not a concession to inclusivity. It is an epistemic requirement for seeing the system accurately.

Together these four patterns constitute the cognitive foundation for systems mapping, the act of making the system visible in a form that allows its parts, relationships, boundaries, and embedded perspectives to be examined honestly and collectively.[17] Making the system visible before reaching for a solution is not a preliminary step on the way to the real work. It is the real work.[17][18] This paper is, in one sense, a partial map of a system. It does not resolve the wicked problem of the Learning Health System. It attempts to make that problem more visible, more precisely named, and more honestly held, in the conviction that a system cannot be improved by agents who cannot see it clearly.

VII: Building the Ecosystem

This paper has traced a specific arc. It began with a conversation, with the recognition that a system described as healthcare has organized itself primarily around sick care, and that a system capable of learning from its own practice toward the goal of genuine health remains largely unbuilt. It named that gap as a wicked problem, structurally resistant to the kinds of solutions that work on complicated problems. It introduced a thinking design lens, VMCL, that reveals where and why the organizational design of American healthcare has been misaligned with a learning mission. It examined three conditions, EHR burden, payer interference, and physician burnout, not as a comprehensive catalogue of everything wrong but as a coherent illustration of a system doing exactly what it was designed to do, which is the wrong thing. And it argued that before solutions can be designed, the system must be mapped, using the cognitive tools of Distinctions, Systems, Relationships, and Perspectives, so that what is actually happening can be seen clearly by the people responsible for changing it.

What comes next is not a conclusion in the conventional sense, because wicked problems do not conclude. They develop. They yield to sustained, cross-disciplinary, honest engagement over time, or they do not yield at all. And that engagement, to be genuine, cannot be organized as a committee or delegated to a working group. It has to function as an ecosystem.

An ecosystem, in the organizational sense, is not simply a collection of stakeholders. It is a community of interdependent actors whose collective behavior produces outcomes that no single actor could generate alone, and whose health depends on the health of every part. The Learning Health System cannot be built by clinicians alone, or technologists alone, or policymakers alone, or systems thinkers alone, because each of those communities has a partial view of the system, and partial views applied with confidence have contributed to the problem as much as to any solution. What the Learning Health System requires is an ecosystem response, one in which diverse and genuinely interdependent actors develop a shared sense of responsibility for the knowledge the system is capable of generating and for the patients whose outcomes depend on whether that knowledge is actually used.

Several conditions define what a functional ecosystem for this work looks like.

Patients must be active contributors, not symbolic participants. The Stanford course materials that informed this paper make a point worth stating directly: in the Learning Health System, every patient is also a research participant, and their data represent an opportunity to learn.[11] The ethical framework developed by Ruth Faden, Nancy Kass, and their colleagues[25] argues that patients have not only rights but obligations within a learning health system, specifically an obligation to contribute to the knowledge that the system generates for their benefit and for the benefit of others, particularly when the risk to them is minimal. Designing health systems that honor that relationship, rather than treating patients as subjects to be protected from the learning process, is one of the most important organizational design challenges the field faces.

Health system leaders must be willing to ask honest questions about what their organizations are actually producing. The wicked problem of the Learning Health System will not be solved by a consultant engagement, a technology platform, or a strategic planning cycle. It will be addressed, partially and incrementally, by leaders who are willing to hold the discomfort of answers that do not reflect well on past choices and design differently in response to what they discover. That requires vision that is genuinely about learning and patient outcomes. It requires mission in the form of simple rules that every agent in the organization can carry and enact. It requires capacity built and aligned for the right purpose. And it requires learning mechanisms that are honest, structural, and actually connected to changed practice.

The ecosystem must also have a convening architecture. Calling for cross-disciplinary engagement on a wicked problem is easy. Designing the conditions under which that engagement can actually happen is considerably harder. In June 2020, the author designed and led SparkJam 2020, a statewide initiative convened through The Rocket Factory in partnership with Activation Capital, the VCU da Vinci Center for Innovation, and other Virginia-based organizations.[22] The initiative brought together entrepreneurs, technology visionaries, business strategists, and community leaders to collaborate in real time on solutions to challenges facing small businesses during the pandemic. The methodology that made it work rested on a specific structural logic: a small group of influential leaders set the agenda, identified the most consequential problems, and recruited a broader population of participants whose direct knowledge and diverse perspectives were needed to work those problems in depth. Structured sessions generated insights that no individual perspective could have produced alone. The broader group returned its work to the leadership tier for synthesis and prioritization, and working groups carried specific initiatives forward. That architecture, a credible leadership tier, broad and diverse participation, structured synthesis, and sustained working group commitment, is precisely what ecosystem convening for the Learning Health System requires.

This paper is itself a beginning and not an answer. It is a partial map of a system far larger and more complex than any single document can represent. What it hopes to contribute is a quality of framing adequate to the problem’s actual complexity. The ecosystem that the Learning Health System requires is waiting to be convened. The methodology exists. The will to build it is what remains to be found.

VIII: AI Implications — When Upstream Conditions Corrupt Downstream Intelligence

The organizational design argument this paper has been making has urgent implications that extend beyond health system walls and into the ambitions of every health technology company, AI developer, and investor currently betting that data-driven tools will transform American healthcare. The case for cross-disciplinary convening made in Section VII is not merely about improving care delivery. It is also about creating the organizational conditions under which technology can actually function as promised. Because the technology being deployed into American healthcare today is only as trustworthy as the data it learns from. And that data was produced by the system this paper has been describing.

Any health technology company seeking to leverage healthcare data to improve patient outcomes must first understand and reckon with what is happening upstream of that data. The organizational conditions under which data is generated determine what that data actually contains. This is not a theoretical concern. It is an engineering one, with direct consequences for patient safety.

Machine learning models learn from the data they are given. They do not evaluate the conditions under which that data was produced. They do not know whether the physician who entered a clinical note was on hour eleven of a shift, copying and pasting from a prior visit to manage an impossible documentation burden, or making a fully considered clinical judgment after a thorough examination. They do not know whether a treatment decision reflected the best available evidence or the path of least resistance through a prior authorization process. They do not know whether a diagnostic code was selected because it most accurately described the patient’s condition or because it was the code most likely to be reimbursed. The model sees the data. It cannot see the system that produced it. That is the job of the humans who build and deploy these tools. And it is a job that is not yet being done with sufficient rigor or honesty in the current wave of enthusiasm for AI in healthcare.

A well-known illustration in machine learning circles, included in the Stanford AI for Healthcare coursework that is part of this author’s ongoing study,[31] captures the failure mode precisely. During the Cold War, the US military hired computer scientists to develop a model that could identify Russian tanks in photographs. The model performed perfectly on the test set. In a live field test it failed completely, performing worse than random guessing. The reason: Russian tank photographs had been taken in winter conditions and American tank photographs in summer conditions. The model had not learned to identify tanks. It had learned to identify weather. It was, in the precise technical sense, a weather classifier dressed as a tank detector.[31]

The same failure mode has been documented in clinical settings. A machine learning model developed to detect pneumonia from chest X-rays outperformed human radiologists in controlled testing. In a small clinical deployment it failed. The model had learned to use the L marker, a physical positioning marker visible in the X-ray images, as a signal to distinguish between the two hospital systems in its training data. One hospital had a one percent prevalence of pneumonia. The other had a 34 percent prevalence. The model did not need to read the X-ray clinically. It learned to read the marker institutionally, and used that artifact rather than any clinical feature to predict pneumonia.[31] It was not learning medicine. It was learning to tell the hospitals apart.

These failures share a common structure. In each case the model learned the wrong signal because the training data encoded something other than the clinical reality the model was supposed to capture. The model was not broken. The data was. And the data was compromised not by random noise but by systematic, directional bias baked into the conditions under which it was produced. This is precisely what the three conditions examined in Section V create for any AI or machine learning system trained on American healthcare data at scale.

It is worth noting that the organizational conditions examined in this paper represent one category of the data bias problem in healthcare AI, and not the only one. The research literature identifies additional sources of bias that compound what has been described here, including the dynamic nature of medical practice over time, which causes historical EHR data to accumulate outdated correlations and effectively expire as a reliable training source as clinical practices evolve, and the demographic non-representativeness of many health system datasets, in which race, ethnicity, gender, and socioeconomic status are inconsistently captured or reported across studies, raising serious questions about whether AI models trained on such data can perform equitably across the full diversity of patients they will ultimately serve.[31]

When 90 percent of clinicians report using copy-paste functionality to manage documentation burden, and when by one estimate 50 percent of the text in a given clinical note is duplicated from prior notes,[27][28][29] the clinical notes that constitute training data for natural language processing models are not accurate records of clinical reasoning. They are records of documentation behavior under pressure. When prior authorization requirements shape which treatments are administered and which are abandoned, the treatment decisions that feed outcome models do not reflect clinical judgment applied to patient need. They reflect what the payer approved. When burned-out physicians experiencing cognitive fatigue make more documentation errors, a connection the research literature supports directly,[30] the signal in the data degrades in direct proportion to the degradation of the workforce producing it.

The research on EHR data quality confirms that these are not marginal concerns. A systematized review published in 2025 examining EHR data quality in critical care settings found that missing data rates exceeded 80 percent for some variables, that EHR-related medication errors comprised 34 percent of all medication errors in ICUs with one-third having life-threatening potential, and that copy-paste prevalence reached 82 percent in residents’ progress notes.[26] The same review found direct and measurable consequences for machine learning: sepsis detection models that achieved strong performance in internal validation dropped significantly in external validation under real-world conditions, a degradation the authors attributed directly to data quality issues pervasive in the underlying EHR data.[26]

The Stanford coursework poses the right question directly: the issue is not whether the data exists. Medical data now doubles every eight to twelve months and there is more of it than ever before. The better question is whether that data is actually usable for the intended purpose.[31] In the current organizational state of American healthcare, the honest answer is not exactly.

This does not mean AI has no role in healthcare. It means the role AI can play is constrained and shaped by the organizational conditions that produced the data it learns from. A 2025 perspective published in npj Health Systems argues precisely this point, noting that while the LHS ecosystem has been well described and its potential widely endorsed, operationalizing the LHS in the era of artificial intelligence requires deliberate attention to data governance, workforce development, and institutional design, the same organizational prerequisites this paper has been examining.[14] The organizational design work this paper has been describing, building genuine Learning Health Systems with aligned vision, mission, capacity, and learning functions, is not merely a clinical improvement agenda. It is the prerequisite for trustworthy AI deployment in healthcare. A health system that has not addressed the upstream conditions producing biased data cannot deploy AI safely or effectively. It will automate the distortions already present in its data and present the result as intelligence. Health technology companies that build on that foundation without looking upstream are not just taking a technical risk. They are taking a patient safety risk. And they are building businesses on data they do not fully understand.

IX: Strategic Implications — The Cost of Not Learning

This paper has operated at two levels simultaneously, and it is worth naming that distinction clearly before drawing it to a close. At the macro level, the Learning Health System is a vision for what American healthcare as a sector could become: a system in which knowledge generation is so embedded in the delivery of care that improvement becomes continuous, self-reinforcing, and oriented genuinely toward the people the system exists to serve. At the micro level, it is an organizational design challenge that must be addressed institution by institution, health system by health system, through specific and deliberate choices about vision, mission, capacity, and learning. The wicked problem lives at the macro level. The work of addressing it happens at the micro level. And the cost of not doing that work accumulates at both levels simultaneously, in individual clinical encounters that produce biased data, in technology deployments built on compromised foundations, in physicians who leave the profession, and in patients who do not receive the care the system was capable of providing if it had been designed to learn.

Gil Bashe argued that American healthcare is not failing for lack of innovation, investment, or talent. It is failing because it has lost sight of the people it exists to serve.[2] This paper has tried to show that losing sight of people and losing the organizational capacity to learn are not two separate failures. They are the same failure, expressed differently depending on where you are standing in the system. The burned-out physician who copies and pastes a clinical note at the end of an impossible shift has not lost sight of their patients. The system that created those conditions has. The EHR that generates data optimized for billing rather than clinical fidelity has not lost sight of patients. The design decisions that produced it have. The AI model that learns the wrong signal from compromised training data has not failed its patients. The upstream conditions that corrupted the data before it ever reached the model have.

The cost of not learning is not abstract. It is clinical. It is financial. It is technological. And it is human. At the macro level it is a sector that has spent nearly two decades describing a vision of continuous learning and improvement while building the organizational conditions that make that vision structurally unreachable. At the micro level it is every health system that has adopted the label of a Learning Health System without asking honestly whether its vision is felt, its mission is enacted, its capacity is aligned, and its learning loops actually close. The gap between those two things, between what is said and what is designed, is where patients fall through.

This paper has not proposed a solution. It has drawn a map. The map shows a system doing exactly what it was designed to do, which is the wrong thing, and it names the organizational thinking, the VMCL lens, the DSRP cognitive tools, the systems mapping discipline, that would allow leaders at every level to see that clearly and begin designing differently. It has also named what is at stake for those who choose not to look. For health system leaders the cost of not learning is an organization that optimizes toward the wrong destination and calls it excellence. For policymakers the cost is interventions that address symptoms without touching causes. For health technology companies the cost is products built on data they do not understand, deployed into systems they have not mapped, producing outcomes they cannot fully explain or defend. And for patients the cost is a system that was capable of learning how to serve them better and chose, through a thousand organizational design decisions made without that possibility in mind, not to.

The Learning Health System is not an idea whose time has not yet come. It is an idea whose organizational prerequisites have not yet been built. Building them is the work. It is hard, sustained, cross-disciplinary, and uncomfortable. It requires the kind of thinking this paper has been describing: honest, structural, willing to see the system as it is rather than as its mission statements describe it. It requires leaders at the macro level of American healthcare policy and at the micro level of every individual health system who are willing to ask whether they are designing for learning or designing for something else and calling it learning.

The conversation is open. The map is incomplete. The cost of not continuing it is borne by patients. That is reason enough to begin.

Citations

[1] Olsen, L.A., Aisner, D., and McGinnis, J.M., editors. Institute of Medicine (US) Roundtable on Evidence-Based Medicine. The Learning Healthcare System: Workshop Summary. Washington, DC: National Academies Press, 2007. PMID: 21452449. DOI: 10.17226/11903. Available at: https://pubmed.ncbi.nlm.nih.gov/21452449/ and https://www.ncbi.nlm.nih.gov/books/NBK53494/

[2] Bashe, Gil. Healing the Sick Care System: Why People Matter. Thought Leader Press, February 1, 2026. https://www.amazon.com/Healing-Sick-Care-System-People/dp/1613431805

[3] Cabrera, Derek and Laura Cabrera. Flock Not Clock: Design, Align, and Lead to Achieve Your Vision. Plectica LLC, 2018. ISBN: 978-1948486019. https://www.amazon.com/FLOCK-NOT-CLOCK-DESIGN-ACHIEVE-ebook/dp/B07DFPWTDS

[4] Cabrera Research Lab. VMCL Overview. Cabrera Research Lab Blog. https://www.cabreralab.science/blog/categories/vmcl

[5] Rittel, Horst W.J. and Melvin M. Webber. “Dilemmas in a General Theory of Planning.” Policy Sciences, vol. 4, 1973, pp. 155-169.

[6] Grewatsch, Sylvia, Steve Kennedy, and Pratima Bansal. “Tackling Wicked Problems in Strategic Management with Systems Thinking.” Strategic Organization, 2023. https://journals.sagepub.com/doi/10.1177/14761270211038635

[7] Dr. Lorna Breen Heroes’ Foundation. “Burnout.” https://drlornabreen.org/burnout/

[8] The Physicians Foundation. “2022 Survey of America’s Physicians.” https://physiciansfoundation.org/press-releases/npsa-day-2022/

[9] American Medical Association. “2024 AMA Prior Authorization Physician Survey.” https://www.ama-assn.org/system/files/prior-authorization-survey.pdf

[10] “Usability Challenges in Electronic Health Records: Impact on Documentation Burden and Clinical Workflow: A Scoping Review.” Journal of Evaluation in Clinical Practice, 2025. https://onlinelibrary.wiley.com/doi/full/10.1111/jep.70189

[11] Stanford University School of Medicine. Course materials on Learning Health Systems and research ethics. Materials on file with author.

[12] Cabrera Research Lab. “Simple Rules.” Cabrera Research Lab Glossary. https://help.cabreraresearch.org/simple-rules

[13] Cabrera Research Lab. “Complex Adaptive System (CAS).” Cabrera Research Lab Glossary. https://help.cabreraresearch.org/complex-adaptive-system-cas

[14] Steel, Peter A.D., Gabriel Wardi, Robert A. Harrington, and Christopher A. Longhurst et al. “Learning health system strategies in the AI era.” npj Health Systems, vol. 2, article 21, June 17, 2025. https://www.nature.com/articles/s44401-025-00029-0

[15] Tenenbaum, J.D. et al. “Accelerating a learning public health system: Opportunities, obstacles, and a call to action.” Learning Health Systems, 2024. https://onlinelibrary.wiley.com/doi/10.1002/lrh2.10449

[16] “Implementing the learning health system paradigm within academic health centers.” Learning Health Systems, 2023. https://pmc.ncbi.nlm.nih.gov/articles/PMC10797573/

[17] Cabrera, D., Cabrera, L. “Why You Should Map: The Science Behind Visual Mapping.” White paper. Cabrera Research Lab, New York, 2018. https://www.researchgate.net/publication/349868707_Why_You_Should_Map_the_science_behind_visual_mapping

[18] Cabrera, L. and Cabrera, D. “Adaptive Leadership for Agile Organizations.” In Cabrera, D., Cabrera, L. and Midgley, G. (Eds.), Routledge Handbook of Systems Thinking. Routledge, London, UK, 2021. Draft preprint on file with author.

[19] Cabrera, Derek. “Distinctions, Systems, Relationships, and Perspectives (DSRP): A Theory of Thinking and of Things.” Evaluation and Program Planning, vol. 31, no. 3, 2008, pp. 311-317. https://pubmed.ncbi.nlm.nih.gov/18554716/

[20] Cabrera, Derek and Laura Cabrera. “DSRP Theory: A Primer.” Systems, vol. 10, no. 2, 2022. https://www.mdpi.com/2079-8954/10/2/26

[21] Cabrera Research Lab. “The Four Simple Rules of Systems Thinking: The Distinction Rule.” Cabrera Research Lab Blog, cabreralab.science. Available at: https://www.cabreralab.science/post/the-four-simple-rules-of-systems-thinking-the-distinction-rule

[22] The Rocket Factory. “The Rocket Factory Presents SparkJam 2020 to Benefit the Virginia 30 Day Fund.” PR.com, June 2020. https://www.pr.com/press-release/814285

[23] U.S. Department of Health and Human Services. “HITECH Act Enforcement Interim Final Rule.” Health Information Technology for Economic and Clinical Health Act, enacted as part of the American Recovery and Reinvestment Act of 2009, Public Law 111-5. Available at: https://www.hhs.gov/hipaa/for-professionals/special-topics/hitech-act-enforcement-interim-final-rule/index.html

[24] Rotenstein, L.S. et al. “System-Level Factors and Time Spent on Electronic Health Records by Primary Care Physicians.” JAMA Network Open, 2023. PMC: https://pmc.ncbi.nlm.nih.gov/articles/PMC10665969/

[25] Faden, Ruth R., Nancy E. Kass, Steven N. Goodman, Peter Pronovost, Sean Tunis, and Tom L. Beauchamp. “An Ethics Framework for a Learning Health Care System: A Departure from Traditional Research Ethics and Clinical Ethics.” Hastings Center Report, Special Issue, January-February 2013, pp. S16-S27. DOI: 10.1002/hast.134. PubMed PMID: 23315888. Available at: https://pubmed.ncbi.nlm.nih.gov/23315888/

[26] “Discovery of data quality issues in electronic health records: profound consequences for critical care medicine applications — a systematized review.” PMC, 2025. https://pmc.ncbi.nlm.nih.gov/articles/PMC12784561/

[27] Tsou, A.Y. et al. “Safe Practices for Copy and Paste in the EHR: Systematic Review, Recommendations, and Novel Model for Health IT Collaboration.” Applied Clinical Informatics, 2017. https://pubmed.ncbi.nlm.nih.gov/28830856/

[28] Urology Times. “Why is copying and pasting in the EHR such a problem?” February 2026. https://www.urologytimes.com/view/why-is-copying-and-pasting-in-the-ehr-such-a-problem-

[29] AMA Journal of Ethics. “How to Teach Good EHR Documentation and Deflate Bloated Chart Notes.” November 2025. https://journalofethics.ama-assn.org/article/how-teach-good-ehr-documentation-and-deflate-bloated-chart-notes/2025-11

[30] “Burnout Related to Electronic Health Record Use in Primary Care.” PMC, 2023. https://pmc.ncbi.nlm.nih.gov/articles/PMC10134123/ [31] Stanford University School of Medicine. Course materials: Fundamentals of Machine Learning for Healthcare. Lecture transcripts on data bias, the Russian tank problem, clinical machine learning applications, medical data shelf life, and demographic representativeness in EHR-based AI research. Part of the AI for

The post Garbage In, Garbage Out: The Organizational Crisis Beneath Healthcare’s AI Gold Rush appeared first on Medika Life.

Health systems are not stepping into artificial intelligence from a place of calm or control. In the United States, spending now exceeds $4.5 trillion, with a significant share tied up in administrative work that adds complexity more than clarity. Clinicians are caring for more patients, navigating more data and making more decisions under pressure than ever before. The system is stretched. Artificial intelligence is entering at a moment when change is no longer a choice.

The discussion drew on the experience of three leaders who are not observing this shift. They are guiding it. Hal Wolf leads HIMSS, influencing digital health policy and implementation across more than 100 countries. Isaac Kohane, MD, PhD, Chair of Biomedical Informatics at Harvard Medical School, has spent four decades defining how data informs clinical care. Ran Balicer, MD, Chief Innovation Officer at Clalit Health Services, operates within one of the world’s most integrated health systems, where data and care are aligned across generations.

These are not just star panelists. They are system-wide architects.  What emerged from the hour-long conversation was not what artificial intelligence can do. It was a recognition that it is already doing more than most systems are prepared to guide and govern.

Dr. Kohane captured the tension immediately. “I think that we have to worry about the fact that we’re going both too slow and too fast.”

That statement reflects a reality many leaders feel but rarely express. Governance takes time because it must. Patient safety, validation and accountability require structure. Practice moves in real time. Clinicians do not have the luxury of waiting for perfect systems.

“They’re so desperate to do right by their patients to use other resources,” Dr. Kohane adds.

That instinct is not a weakness. It reflects a commitment to doing what is right for the patient. When clinicians turn to external AI tools, they are seeking clarity, speed, and confidence in their decisions. Artificial intelligence is already present at the point of care, shaping how physicians assess information, validate thinking, and move forward. The system is not adopting AI. The system is catching up.

This creates a condition that is difficult to measure and even harder to manage. Different clinicians use different ChatGPT platforms. Those tools produce different answers. Different assumptions shape those answers. Over time, consistency erodes. The system begins to operate with multiple definitions of truth (and the risk of varied outcomes).

Dr. Kohane’s warning is not about misuse. It is about misguided permanence. “The worst outcome will be if the worst parts of medicine get concrete poured over it, by AI.”

Artificial intelligence does not fix a system; without leadership, it accelerates the integration of incorrect assumptions. If workflows are inefficient, they become more efficiently inefficient. If bias exists in data, it becomes more precise. If fragmentation defines care, it scales.

This is not a failure of technology. It is a mirror held up to system-wide leadership.

Hal Wolf, among the health sector’s leading policy and operational voices, grounded this moment in proven experience. Health care has seen this pattern before. When internet connectivity entered hospitals, clinicians moved faster than governance. They created access where it was needed. Systems responded later. Risks were discovered after adoption.

Artificial intelligence now follows that same trajectory, though at far greater speed and with far greater consequences. Web connectivity gave quick access to information. Artificial intelligence influences how that information is interpreted and acted upon.

“We have to go faster,” Mr. Wolf said. “But there needs to be structure around it.”

That is the leadership challenge of this moment. Speed without structure creates exposure. Structure without speed creates irrelevance. The tension between the two is not something to resolve. It is something to manage continuously.

The industry has predictably responded to artificial intelligence. It has started where risk is lowest and return is clearest. Documentation, scheduling and revenue cycle optimization have become the entry points. These applications reduce burden and improve efficiency. They are necessary. However, they are not transformational.

The shift occurs when artificial intelligence moves into clinical decision-making. At that point, the question is no longer whether the system works. The question becomes whether it should be trusted.

Who owns a decision informed by an algorithm? How is accuracy validated? What happens when a clinician disagrees with a recommendation? These are not technical questions. They are questions of accountability. Artificial intelligence does not assume responsibility. It does not carry consequence. That remains with leadership.

Dr. Balicer reframed the conversation, shifting how the room thought about artificial intelligence. “There’s no such thing as AI neutrality. Algorithms are just opinions embedded in code.”

That insight is easy to acknowledge and difficult to operationalize. Every model reflects choices. What data is included? What outcomes are prioritized? What trade-offs are accepted? Those decisions are embedded in the system, shaping how it interprets information.

When a health system adopts an AI tool, it is not simply implementing technology. It is adopting a perspective.

At Clalit Health Services, alignment across payer and provider creates a system where priorities are consistent. Even there, external AI models introduce new assumptions. Those assumptions may not align with the system’s goals. If leadership does not define its own values, it inherits someone else’s.

This becomes real in proactive care. Artificial intelligence enables systems to identify patients at risk before they present. It allows for earlier intervention, often improving outcomes.

It also creates a new kind of pressure. “The toughest choice is what not to do,” Dr. Balicer said.

That statement deserves more attention than it receives. Health care has been built around responding to need. Artificial intelligence introduces the ability to anticipate it. When every patient can be flagged, every risk predicted and every intervention suggested, the system is no longer constrained by insight. It is constrained by capacity.

Artificial intelligence expands what can be done. It does not expand who can do it. Leadership becomes the act of choosing who does what based on validated data.

There is a moment that captures this shift. Imagine a primary care physician starting the day not with a schedule of patients who have called for appointments, but with a list generated by AI identifying individuals who are likely to experience clinical complications in the next six months. Some will develop chronic conditions. Some will require hospitalization. Some can be helped now – preventively.

The physician cannot see them all. Artificial intelligence expands what is possible. Leadership decides what is essential and permissible.

The industry often responds to complexity with activity. Organizations pilot, test and explore. They engage broadly without committing deeply. This creates motion. It rarely creates progress. Pilots are nothing more than experiments. At some point, leadership must decide what to scale, what to stop and what defines value.

Hal Wolf grounded the conversation in discipline. Without a defined, shared objective, effort becomes noise. Pilots create learning, though they often avoid decision-making. Leadership requires clarity. What problem are we solving? What outcome defines success? What are we willing to prioritize? Without those answers, artificial intelligence adds another layer of complexity to an already complex system.

Dr. Kohane brought the conversation back to the discipline of leadership. It cannot remain abstract. It must be informed by experience.

“Go and pay a few bucks and use three or four of the models… get a feel for what this does,” Dr. Kohane advised.

That is not a call for technical fluency. It is a call for leadership proximity. Leaders cannot guide what they do not understand. Artificial intelligence does not behave consistently across models. It produces different answers, shaped by different assumptions. Without direct engagement, those differences remain hidden, and leadership becomes removed from the very decisions it is responsible for guiding.

This is where many organizations hesitate. Artificial intelligence feels complex and complexity invites delegation. At this moment, delegation creates distance. Leadership is required to move closer, not further away.

Artificial intelligence is not reducing the role of leadership. It is redefining it.

This is not a gradual transition. It is already underway. Artificial intelligence is embedded in workflows, shaping decisions and influencing behavior in real time. The system is adapting whether leadership is ready or not.

The question is no longer whether artificial intelligence will shape the future of health. It will. The question is whether leadership will shape how it is applied.

Artificial intelligence will not fix health. It will scale whatever we allow it to touch. The question is whether it will scale what is best in health or what we have yet to fix.

The post AI Will Not Fix Health Care – Leadership Might appeared first on Medika Life.

But from the perspective of philosophers of science, when do such shifts in scientific approaches actually occur?

Thomas Kuhn’s Perspective

Kuhn believed that changes in scientific approaches resemble political revolutions. Simply put, when a government can no longer manage society or effectively administer its affairs, dissatisfaction gradually spreads among the public and opposition begins to form. In other words, the inability to respond to society’s needs becomes the driving force behind revolutionary movements. This process continues until a capable system emerges that can meet those needs, eventually leading to the establishment of a new order.

A similar process occurs in what Kuhn calls scientific revolutions. According to him, in every era the majority of scientists accept and follow a general framework. Kuhn refers to this dominant framework — which contains a collection of theories and practical models — as a paradigm. Paradigms are patterns widely followed by scholars, such as the paradigm of modernity or the paradigm of cognitive science.

As long as these paradigms remain aligned with the requirements of life and are capable of addressing existing problems, they continue to be valued and are used in major policy frameworks. However, when a dominant paradigm fails to respond to contemporary challenges and the solutions derived from it prove ineffective at addressing large-scale needs, doubts arise about its continued relevance. Under such circumstances, dissatisfaction intensifies to the point that scholars begin to consider laying the groundwork for a new, updated paradigm.

In his book The Structure of Scientific Revolutions, Kuhn emphasizes that scientific transformations are not linear or step-by-step processes. Rather, they are complex and revolutionary developments in which social and historical factors play a crucial role. Under normal conditions, scientists operate within the framework of an accepted paradigm — what Kuhn calls normal science. However, when persistent anomalies emerge and the paradigm proves incapable of addressing them, the existing structure eventually collapses and a scientific revolution occurs.

Karl Popper’s Theory of Science

Like many philosophers of science, Popper believed that change is not only inevitable but also a necessity. The Popperian view rests on the principle of falsifiability. In this framework, science begins with a problem, and solving a problem means finding solutions to existing challenges. As long as a scientific theory remains open to criticism and falsification, it retains the capacity to address and solve problems.

In Popper’s view, bold conjectures do not weaken science; rather, they strengthen it. Solutions proposed under the principle of falsifiability help correct previous errors, and this is precisely where the strength of the scientific approach lies. If existing approaches are not falsifiable, they lose the possibility of logical trial and error and are therefore considered weak. In such cases, the need for a shift in approach and the introduction of new models becomes evident.

Popper believed that learning is essentially problem-solving guided by the principle of falsifiability.

To move beyond temporary and ineffective solutions, followers of science must avoid false certainties, accept falsification, and search for effective alternatives.

The Need to Shift from Data-Driven AI to Learning-Science-Based AI

Today, numerous criticisms are directed at the purely computational and mechanical approach to artificial intelligence. In constructive critiques, the goal is not to deny the existence of large language models; rather, the central question concerns how and under what conditions they should be used. There is a growing consensus that the closer artificial intelligence moves toward the essence of human cognition, the lower its potential risks become.

In recent years, I have repeatedly emphasized that human theories and perspectives must be reexamined through a technological and contemporary lens so that the nature of the human mind is properly reflected in technologies that themselves were modeled after it. 

My focus lies on deep theories of learning (including cognitive approaches, neuroscience, behaviorism, evolutionary perspectives, structuralism, and other related frameworks).

In this direction, the following steps appear essential:

1. Integrating human and computational perspectives

The current approach, which relies excessively on probability laws in large language models, must be integrated with psychological perspectives. A reasonable solution is to pursue interdisciplinary studies and systematic research in this area.

2. Revisiting theories of the learning sciences

Theories that analyze the human mind and behavior should be reassessed by specialists, and their practical dimensions should be extracted for application in advanced technologies.

3. Developing integrative (hybrid) approaches

Experts should develop comprehensive perspectives on learning derived from multiple scientific approaches so that, based on research rather than mere speculation, practical recommendations can be provided to designers and engineers.

In general, the time has come to move beyond a purely logical and mathematical approach toward a human-centered perspective. To address the concerns and challenges surrounding artificial intelligence, we must return to systematic and interdisciplinary research.

The era of relying on personal opinions without a research foundation — or on mathematical rules alone — has come to an end. Now is the time to revisit the learning sciences from a new perspective in order to realize truly human-centered artificial intelligence

Author’s Note:

The ideas presented in this article are part of a broader research project. I am currently working on a comprehensive book on a new approach to human-centered artificial intelligence with a strong emphasis on the learning sciences. While a detailed and systematic discussion of these concepts is presented in Chapter Two, the book also includes a dedicated chapter introducing the new paradigm’s framework. Furthermore, at least one chapter is specifically focused on the practical methods and applied implications of this approach for implementation in artificial intelligence systems.

References

• Kuhn, T. S. (1962). The Structure of Scientific Revolutions. University of Chicago Press.

• Popper, K. (1959). The Logic of Scientific Discovery. Hutchinson.

• Popper, K. (1963). Conjectures and Refutations: The Growth of Scientific Knowledge. Routledge.

The post The Shift from Pure Modernity to Human-Centered Modernity appeared first on Medika Life.