Every, the tech publication, asked twenty five founders, designers, investors, and writers to each stake one claim about work after automation. Short essays. One thesis apiece. Read one and you get a startup op-ed. Read all twenty five and something better happens: you get a map of where smart people outside medicine think the human/machine boundary actually sits.
I care about that boundary for a specific reason. My clinical epistemology work rests on one wager: the explicit portion of clinical reasoning can be written down as steps and gates and enforced on machines. What data does this question need. Do we actually have it. Does it mean what it appears to mean. Is it sufficient. That wager has an obvious complement. If the explicit part can be gated, what exactly is the part that cannot? These essays are twenty five answers to that question. (They were all answering the same assigned prompt, so the convergence is partly the assignment. Where they converge is still instructive.)
And in the same week I compiled this, the FDA opened a public docket asking how it should test generative AI medical devices. The agency's framing question is not "how accurate is it on a leaderboard." It is "does the machine know what it knows." Same boundary. Different building.
The essays sort themselves into five clusters. Four of them describe where automatable thought ends. The fifth is not a boundary at all. It is a prescription for what to build at the edge.
Dan Shipper puts it most cleanly: AI commoditizes the residue of human expertise, meaning whatever can be made explicit enough to train on. Models know what has been done. The human in the room knows what is needed right now, with this patient, this dataset, this conversation. Paul Millerd extends the point: automation reaches job shaped work, the subset formalized enough to price, and never the illegible remainder (care, presence, meaning).
Yash Tekriwal runs the argument the other way. As agents absorb execution, the durable human skill is more explicitness, not less. Writing instructions for machines is an audit of your own thinking. Both directions are right, and my frameworks need both. You make the reasoning explicit precisely to find out where the explicit part ends.
Sari Azout draws the sharpest line in the series. AI thrives where outcomes are verifiable. The decisions that matter rarely are. Her formulation: AI can tell us what is probable, but not what is worth wanting. Nir Zicherman makes the same cut operationally (coding automates because outputs are checkable; ambiguous work does not). Jim Prosser adds a twist I keep thinking about: what survives automation is work where the customer can actually perceive the difference between adequate and excellent. Much of medicine fails that perceptibility test at the bedside, which is exactly why credentialing, peer review, and tiered authority exist.
This is the is/ought line, and clinical decision analysis has known it for decades. It also lands squarely on a gap my own frameworks have already confessed in writing: confidence without a values component means every recommendation asserts an unwritten utility function. The machine supplies the probability. Someone still has to supply the worth.
Anne-Laure Le Cunff: when answers become abundant, value migrates to deciding what deserves attention and knowing when your starting question was wrong. Dan Pupius adds the decision theoretic version: stop predicting, classify your bets as reversible or irreversible, and optimize how fast you notice you are wrong. That maps directly onto tiered authority. An observation is a reversible act. An actionable recommendation is much less so. The authority a system earns should scale with the reversibility of acting on it.
Joe Hudson defines wisdom as the residue of mistakes, metabolized by time and reflection. Khe Hy points out that AI's competence ends at the edge of the recorded world (the sidebar conversation after the panel never enters any dataset). The two essays that matter most for medicine are about formation. Willem Van Lancker argues that friction is not inefficiency; struggle is how judgment and taste get built, and AI removes struggle by default. Shoshana Berger makes it generational: mastery is earned through mentored failure, and if masters stop teaching apprentices, everything stops learning. Including the AI, which trains on what masters made.
Translate that to a teaching hospital and it stops being philosophy. If AI absorbs the first read of the labs and the first draft of the differential, it absorbs the productive friction by which residents become the attendings whose judgment the whole system depends on. Tier policy may need to be partly pedagogical. That is a workforce question, not a sentimental one.
Karri Saarinen: the slippery feeling of AI products traces to the interface, not the model. The models improve on their own; the harder work is the structure around them. Tina He: agents are ruthless rational actors, and the durable moat is the orchestration and rule enforcement layer. Tom Critchlow builds a shared organizational clock and then bounds his own idea with the best sentence in the series: knowing the time is not the same as knowing what the moment requires. Noah Brier: the insidious failure mode of agentic engineering is not buggy code, it is coherent looking misalignment with intent.
Five people who have never seen my enterprise framework independently restated its thesis. Foundation models commoditize. The value is the governed layer that binds questions to decisions. Trustworthiness is not a property the model emits. It is a property you build around it.
On August 18 the FDA published a discussion paper on regulating generative AI enabled medical devices and opened docket FDA-2026-N-7874 for comment through October 19. Not draft guidance. Not policy. A request for feedback, sitting in a lineage that includes the 2019 AI/ML discussion paper, the GMLP principles, and the 2024 predetermined change control guidance.
The organizing move is a clinician licensure analogy. Regulators do not anticipate every situation a physician will face in a career. They test underlying knowledge, observe performance, and require continuing oversight. The FDA asks whether a version of that model can work for devices. (Illustrative, not jurisdictional. The FDA regulates products, not practitioners.)
The proposed exam is more specific than I expected. Does the device know what it does not know. Does it refuse to answer outside its scope. Does it hand off when it should. Are its numbers right. Does it perform consistently across patient subgroups, including nonstandard dialects and lower health literacy. Does it withstand a prompt written to break it. And will anyone notice when a silent update makes it worse. Calibration, deferral, and scope adherence as first class test requirements. That is epistemic humility written into a regulatory framework.
The gaps are as instructive as the framework. The benchmarks that would test uncertainty handling and robustness mostly do not exist yet (a 2026 audit found roughly nine in ten medical LLM benchmarks never evaluate uncertainty handling at all). The comparator is unstable, since trials keep finding the model alone performing at or above clinicians using it. And a physician who fails oversight faces licensure consequences, while a failed device gets a software update. Competency without consequence is a certificate.
Still, the direction matters. A regulator, an academic liability taxonomy in Nature, an industry agentic maturity matrix, and my own tier ladder keep reaching for the same shape: graded authority instead of binary approval. Different legal objects, different units, and nobody copied anybody. When four parties working on different problems keep drawing the same ladder, the ladder is probably load bearing.
We make clinical thought process explicit, gated, and tiered precisely so that what cannot be made explicit (the question, the values, the felt sense, the accountable commitment) is exercised by humans deliberately and visibly, rather than by accident or not at all.
Sources
Essay series: Every, "Thesis Statements" (After automation), 25 essays. every.to/thesis-statements
FDA paper: FDA CDRH, Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback (Aug 18, 2026). Docket FDA-2026-N-7874, comments close Oct 19, 2026. fda.gov landing page · PDF
WAiR deep dive: FDA GenAI competency review (2026-08-20). wair.ajwein.com/fda-genai-competency-2026-08-20-v1
Benchmark audit: Chen W, et al. Beyond the Leaderboard. arXiv:2508.04325
Liability taxonomy: Lam K, Duffourc M, Sun A, Topol E, Qiu J. Tackling the blurring lines between physicians and AI. Nature 655:1129–1132 (July 2026).
Comparator trials: Goh E, et al. JAMA Netw Open 2024;7(10):e2440969; companion Nat Med 2025 citation under verification.
Full synthesis document: After_Automation_Epistemology_Synthesis_2026-08-24.md, Clinical Epistemology project, _reference folder (includes per essay summaries, URLs, and the two model red team validation note).