Billions of venture dollars are flowing into a single bet: If you can generate a medically sophisticated answer fast enough, you have solved something meaningful in healthcare. The pitch is seductive. Doctors are under extreme time pressure, patients wait months, and large language models (LLMs) can now produce answers that are polished, empathetic, and clinically credible in seconds, at a fraction of the cost.
The problem is that the bet is built on a category error, and medicine may spend the next decade paying for it.
Knowing what matters
The hard part of medicine has never been retrieving information. It is knowing which information matters for this patient, in this moment, under conditions of uncertainty, with incomplete data, real consequences, and constraints that no algorithm has ever had to navigate. The moment a patient’s story does not fit the referral note. The difference between “I’m tired” and “something is very wrong.” The feel of tissue during surgery. The judgment to know when the guideline applies, when it does not, and when the guideline itself is already behind practice.
None of that information has ever been cleanly captured in a database. Much of it never will be.
The industry is confusing medical information with medical judgment. LLMs are extraordinary at synthesizing what has been written down, but much of what makes medicine trustworthy lives somewhere else entirely: in experience, in context, in pattern recognition built over thousands of cases, and in the peer-to-peer clinical reasoning that happens between doctors. That last part is especially important, and it is the part Silicon Valley has most completely ignored.
Clinical judgment is built in the hallway after a difficult case, in the curbside consult, in the “are you seeing this too?” exchange between a cardiologist and an intensivist who would never otherwise cross paths. It’s true that medicine has a uniquely vast formal knowledge infrastructure: think journals, conferences, guidelines, grand rounds and more. But the distributed, real-time, peer-to-peer reasoning layer is where much of clinical intelligence is actually forged, and that layer has always been structurally unprotected.
Scaling reasoning
For a brief and improbable moment, MedTwitter changed that. For all its dysfunction—the pile-ons, the hierarchy games, the performative certainty — MedTwitter gave physicians something medicine had never intentionally built: a real-time, cross-specialty, global clinical commons. A rural emergency physician could post a difficult ECG and hear from an expert within minutes. A trainee could watch senior clinicians debate a study on the same day it was published. A new trial could be challenged, refined, contextualized, and pressure-tested by the people who actually had to take care of patients the next morning. That was MedTwitter’s real value to medicine. It briefly scaled the informal layer of clinical reasoning before collapsing under the incentives of the platform that hosted it.
The failure was also structural. A platform built to maximize attention cannot sustain a community that depends on thoughtfulness, trust, humility, and professional norms. Medicine needs spaces where a doctor can say “I don’t know” or “here’s what we do at my institution” without being punished by an algorithm optimized for outrage and certainty. Doctors need places where disagreement is productive, uncertainty is honest, and where their expertise does not have to perform for patients, journalists, employers, trolls, and strangers all at once.
That absence of a physicians’ commons matters more now than ever. Physicians are navigating the arrival of artificial intelligence into clinical practice without functioning infrastructure for collective interpretation.
What does expertise mean when information retrieval is commoditized? Which tools represent genuine breakthroughs, and which are polished hallucinations? How should a community oncologist evaluate an AI-generated treatment recommendation when the model may have been trained on different patients, in different institutions, with different constraints? How should real-world evidence, local practice patterns, and institutional experience shape the use of these tools?
These questions will not be answered by product demos, glossy benchmarks, or one more AI summary of a paper. The answers will come from rebuilding the distributed, peer-to-peer, real-time clinical reasoning layer that medicine has always depended on but never properly protected.
Integrating information
This is what the replacement narrative gets backward. AI will help patients navigate the system. It will help doctors retrieve information, summarize records, draft notes, flag risks, and support decisions. But as information retrieval becomes cheaper and faster, physicians do not become less important. They move upstream into the harder work of integrating patient-specific context, institutional constraints, lived experience, uncertainty, evidence, values, and risk into decisions that actually affect human lives.
In that world, the most important question doctors ask may not be, “What does the model say?” It may be the older, harder, and more human question: “What would you do?”
Right now, physicians’ collective intelligence is fragmented across siloed Slack channels, private group chats, text threads, and informal back channels. These spaces are often high-trust, but narrow. They do not have the cross-specialty reach, scale, or structure required for true collective sense making. And that experience cannot simply be automated, because making sense of things in medicine is driven by trust, nuance, and credibility. It depends on knowing who is speaking, what they have seen, how they practice, and why their judgment matters. This is the human layer where evidence becomes judgment, and judgment becomes care.
The future of better patient care will not be determined only by what AI knows. It will depend on whether we can unlock the vast knowledge, judgment, and experience that already exists inside physicians and make that collective intelligence available to the people caring for patients in the real world.
The next era of medicine does not need another platform optimized for attention or another tool that treats doctors as endpoints for generated answers. It needs trusted infrastructure for making clinical sense: a place where evidence can be challenged, experience can travel, uncertainty can be discussed honestly, and physicians can help one another decide what knowledge means for the patient in front of them. In the age of AI, that human network is not a retreat from progress. It is the infrastructure that will drive innovation and make progress clinically meaningful.
In our new age of AI, the most important technology may not be the model itself. It may be the community that learns, questions, tests, and ultimately decides what that model means in practice. The future of medicine will belong not only to what machines can know, but to what physicians can discover together.
