The Pragmatist’s AI Manifesto
The Dangers of Premature Pontification
The track record of predicting the long-term trajectory of artificial intelligence is poor. Even prominent experts have watched their confident pronouncements age badly in real time.
In March 2023, Noam Chomsky and his co-authors published a widely read essay in The New York Times arguing—based on GPT-3.5—that large language models faced fundamental, insurmountable limits in reasoning. Less than a week later, GPT-4 launched and performed the very tasks they had declared structurally impossible. Confident claims that neural networks would never write working code, work through multi-step mathematics, or sustain context across complex problems have fared no better. As Google DeepMind CEO Demis Hassabis has noted, predicting ten years out is hard enough in ordinary times; when AI capabilities compound week by week, nobody knows where the ceiling lies. Long-range forecasting is a poor substitute for direct work. And in healthcare—regulated, messy, and stubbornly human—grand predictions rarely survive their first contact with a real clinic.
Yet in classrooms, hospital boardrooms, academic panels, and policy circles, enormous energy still goes into debating speculative third-, fourth-, and fifth-order consequences of AI from the sidelines. I fear that, whilst we need to be cognizant of the risks inherent in this and any technology, the risk is that premature pontification breeds paralysis rather than clarity. Worse, it distracts us from pressing problems that are already here.
While commentators debate what medicine might look like in ten years, the pressures in front of us are immediate:
- Clinicians are burning out on administrative friction—or losing the sense of purpose that brought them into medicine.
- Operational bottlenecks are delaying care for patients who cannot wait.
- Despite historically high spending, millions of people still struggle to access basic services.
- Health systems everywhere face mounting financial strain.
Developing AI Competence
At the Harvard T.H. Chan School of Public Health Advanced Learning Academy, I built our AI in Healthcare curriculum around a simple premise: you develop real judgment in AI the same way you develop clinical judgment—by moving deliberately from understanding to supervised practice to independent action.
In medical education, George Miller’s 1990 Pyramid of Clinical Competence established a fundamental framework for assessing performance. Miller posited that clinicians do not leap straight from textbook theory to bedside mastery. Instead, they progress systematically through four cognitive and behavioral stages: Knows (foundational knowledge), Knows How (applied competence and clinical reasoning), Shows How (guided performance in controlled settings), and Does (independent, real-world action in clinical practice).
Our program’s Learning Pyramid translates Miller’s clinical paradigm directly into the artificial intelligence domain, structuring modern technical capability into three distinct, actionable tiers:
- Base — AI Literacy (Knows): Building a foundational grasp of how AI architectures function, identifying inherent failure modes (such as hallucinations and algorithmic bias), and asking critical, discriminating questions about model reliability before deployment.
- Middle — AI Applications (Knows How & Shows How): Moving from passive understanding to practical execution by evaluating commercial tools, structuring robust prompts, enforcing data privacy guardrails, and managing human-in-the-loop workflows within real healthcare operations.
- Apex — AI Innovation (Does): Achieving true mastery by constructing working software prototypes, deploying custom AI agents, and engineering end-to-end agentic workflows that transform clinical routines.
Ascending this pyramid requires replacing passive commentary with an iterative, experiential learning loop: Ask → Build → Observe → Critique → Revise. Practitioners formulate precise hypotheses (Ask), construct a rapid solution or prompt harness (Build), empirically analyze real output behavior and failure cases (Observe), uncover underlying root causes for systemic discrepancies (Critique), and iteratively refine their design (Revise). Before we can credibly govern, evaluate, or debate the distant future of AI in healthcare, we must engage directly with the work ourselves.
Three practical commitments guide that climb.
1. Learn Intentionally—and Use AI to Study AI
Foundational literacy begins with understanding how a system works before trying to deploy it. That takes structured study, but AI offers an unusual advantage: it is the first general-purpose technology in history that can help teach you how it works.
You cannot ask a steam engine to explain thermodynamics, and you cannot ask an electronic health record to walk you through its database schema. Modern language models and coding agents, by contrast, can serve as patient, interactive tutors alongside a structured curriculum. Rather than relying on second-hand summaries, open a model and test it directly. Ask how it weighs clinical evidence, where a prompt hides unstated assumptions, and why its answers break down when records conflict.
That is the foundation of our core course, Responsible AI for Healthcare: Concepts and Applications, which builds conceptual clarity and practical judgment without a coding background being required or even recommended, really. We also use custom AI learning tools inside the courses themselves so participants can test ideas, inspect mistakes, and get immediate feedback as they learn.
At the same time, honest literacy requires humility. Modern AI is unlike conventional software: it is probabilistic and non-deterministic. At a scale of hundreds of billions of parameters, these models operate at a level of mathematical complexity that no human can trace step by step. For clinicians and scientists trained to expect deterministic rules and clear causal mechanisms, that opacity is uncomfortable—but we have to face it squarely.
In a conversation with Dr. Eric Topol for The AI Revolution in Medicine (co-authored with Dr. Isaac Kohane, Editor-in-Chief of NEJM AI), Microsoft Research President Dr. Peter Lee put it plainly: “We see GPT-4 doing things that we can’t explain given our current understanding of how a neural transformer operates.” Writing in the Hastings Center Report, bioethicist Alex John London made a parallel point for medicine: “Although their designers understand the architecture of these systems… the models themselves can be inscrutable to humans”—much like therapies such as aspirin and lithium, whose clinical usefulness was established decades before their mechanisms were fully understood.
Even if we could map every internal weight of today’s models, the field moves too quickly for static textbook knowledge to hold. Literacy has to be empirical: forming expectations, testing them directly, and learning how these systems actually behave.
2. Apply AI to the Work You Already Know
Moving from Knows to Knows How and Shows How means bringing that literacy into the work you understand best. AI is indifferent to domain, which makes domain expertise more valuable, not less. Whether you practice clinical medicine, run hospital operations, lead public health programs, or manage a health plan, your advantage is knowing where the real workflows break.
Clinicians, researchers, and operators see the daily handoffs, exceptions, and workarounds that outsiders miss. They are in the best position to tell the difference between an administrative task where speed is welcome and a clinical decision where human judgment is non-negotiable.
Start by auditing your own work. Where does administrative friction pull you away from patients, strategy, or research? How do you evaluate an AI vendor’s claim against your actual data, set guardrails around sensitive records, and make sure a human reviewer has the time and evidence to catch mistakes?
This middle tier of the pyramid is the focus of Implementing Healthcare AI into Clinical Practice, where participants work through implementation cases to evaluate tools, redesign workflows, and lead organizational change.
3. Build Something Small and Real
At the top of Miller’s pyramid is Does—solving real problems in practice. I believe that clinicians learn best by building, and the primary objective of our entire suite of courses, working together collectively, is to take clinicians from foundational concepts all the way to acquiring hands-on build literacy: the capability and confidence to turn concrete healthcare operational challenges into functional AI solutions.
You no longer need to be a software engineer to build. Describing a problem clearly, testing a prototype on synthetic data, and watching where it fails triggers the Ask → Build → Observe → Critique → Revise cycle in a way no lecture can replicate. More importantly, building reveals the deployment gap: the distance between a demonstration that looks impressive on a screen and a service that is safe, governed, and dependable in a real clinic.
Our advanced courses are engineered to finalize this trajectory and cement that practical ability:
- Product Innovation (Building AI Solutions to Transform Healthcare): Clinicians apply their cumulative knowledge to define healthcare requirements, architect specifications, and construct fully working AI prototypes from the ground up.
- Operational Innovation (Innovation with AI in Healthcare: Agentic Operations): Clinicians design and build multi-agent operational systems, learning to directly integrate software harnesses, autonomous AI capabilities, and human oversight in live workflows.
You do not need a computer science degree to do this work, but you do need to build something real, compare what happens with what you expected, and let the evidence change your mind.
When you are ready to move from speculation to action—whether by taking a single course, completing three courses for the AI in Healthcare Certificate of Specialization, or pursuing the five-course Advanced Certificate of Specialization in AI in Healthcare Leadership—I invite you to build with us.