The Triangle Offense for AI-Native Medicine

Why physicians need a new posture for AI-native patients: patient context, AI synthesis, and physician judgment inside the same clinical loop. AI-native patients are arriving with structured context, longitudinal data, research, hypotheses, and AI-assisted reasoning. Physicians need collaborative skepticism rather than dismissal. Patients are beginning to arrive with AI-assisted timelines, hypotheses, symptom data, research, and questions. The next clinical posture is not dismissal or surrender. It is collaborative skepticism: taking the patient's work seriously while testing it clinically. AI-native care should move beyond separate patient-AI and physician-AI silos toward a shared working model of the case. Patients are arriving with a second brain Two recent patient stories point toward a future that medicine is not fully prepared for. Not the distant, panel-discussion future where everyone says "transformation" and then goes back to faxing records. The near future. In [Elliot Hershberg's essay about Sid Sijbrandij](https://centuryofbio.com/p/sid), Sid approaches recurrent osteosarcoma with the operating system of a technical founder. After exhausting standard options, he builds a massive information stack around his care: meticulous notes, maximal diagnostics, tissue access, genomic testing, treatment hypotheses, expert conversations, and a personalized therapeutic ladder. The story is extraordinary because Sid is extraordinary. But it is also important because it shows what happens when a patient refuses to remain a passive recipient of fragmented medical information. In [Amy Deng's essay on solving mystery fatigue with AI](https://metalearn.substack.com/p/i-solved-my-mystery-fatigue-with-ai), the pattern is more accessible and more obviously AI-native. She tracks symptoms, nutrition, sleep, menstrual cycle data, labs, blood pressure, and other signals. She uses AI to help decide what to track, analyze the data, generate hypotheses, prepare for clinicians, and identify what additional tests or interventions might be reasonable. Crucially, she also keeps the right safety boundary: risky medical action still requires the care team. These stories are not the same. One is an extreme oncology case involving unusual resources, rare disease, and frontier diagnostics. The other is a patient-led investigation into a chronic, ambiguous symptom. But the shared signal is hard to miss: > Patients are starting to arrive with structured context, longitudinal data, research, hypotheses, and AI-assisted reasoning. > Many physicians are not ready for that. Not because physicians are incapable. Not because patients are always right. Not because AI should be trusted blindly. The problem is more subtle: the traditional clinical posture assumes that the physician is the primary processor of medical information and the patient is the narrator of symptoms. AI changes that relationship. > The patient can now show up with a second brain. That second brain may be flawed. It may hallucinate. It may overfit. It may miss dangerous context. It may make the patient more anxious. But it may also organize months of lived experience better than a ten-minute visit can. It may surface a pattern nobody had time to see. It may help the patient ask a better question. It may turn scattered data into a visit-ready synthesis. It may not have a medical license. It may not even be right. But it is now in the room. > The physician who treats this as a threat will lose the patient. > The physician who treats this as a partner will be better positioned for the next era of care. The old posture is breaking Medicine has long relied on a hierarchical model of decision-making. The patient describes the problem. The physician interprets. The physician orders tests. The physician decides. The patient complies. Clean, tidy, and increasingly fictional. That model was never the whole truth. Good physicians have always listened carefully, respected patient goals, and incorporated patient observations. But the operating model of care still tends to place medical reasoning inside the clinician's head and medical information inside the clinician's record. AI pushes reasoning outward. A motivated patient can now upload labs, organize timelines, compare symptoms against medication changes, create charts, summarize specialist notes, review guidelines, interrogate differential diagnoses, and prepare a focused agenda before the visit. A patient with programming skill can go further: parse exports, correlate wearable data, build personal dashboards, run literature searches, and use AI coding tools to analyze years of data. That does not make the patient the physician. It does make the patient a more capable collaborator. There is some history here. This is not the first time patients have used a new information layer to weaken medicine's monopoly on interpretation. In the nineteenth century, [domestic medicine manuals](https://archives.health.ufl.edu/events/plants-in-art-and-medicine/popular-medical-manuals-in-the-19th-century-united-states/) translated medical knowledge into household language. [Friendly Societies](https://www.massmed.org/About/MMS-Leadership/History/The-Socialization-of-the-Practice-of-Medicine/) let working-class patients collectively contract for care outside elite medical structures. Much later, [AIDS activists](https://embryo.asu.edu/pages/construction-lay-expertise-aids-activism-and-forging-credibility-reform-clinical-trials-1995) became sophisticated lay participants in clinical trial design, access policy, and biomedical knowledge-making. The lesson is not that patient-led medicine is automatically correct. The patent-medicine era is proof that agency can come with real danger. The lesson is that when institutions cannot absorb patient need, patients build workarounds. This is where the physician mindset has to change. The core question is no longer, "How do I stop patients from using AI?" It is, "How do I practice medicine when patients are using AI anyway?" > The answer cannot be dismissal. It also cannot be surrender. Dismissal wastes useful information and alienates engaged patients. Surrender is unsafe because AI systems can produce confident nonsense, miss contraindications, and blur the line between plausible and clinically appropriate. The physician's role becomes more important, not less. But the role changes. > The physician becomes the clinical adjudicator of a larger information loop. The problem is parallel AI silos Most healthcare AI is still being imagined inside the silos medicine already has. We are installing smarter appliances in separate rooms and calling it a connected house. The patient gets a consumer AI tool. They ask it about symptoms, side effects, labs, diet, supplements, medications, and diagnoses. The conversation may be useful. It may help them organize their experience. It may make them more informed. But it often happens outside the clinical context, outside the medical record, and outside the physician's view. The physician gets a different AI tool. It may draft notes, summarize charts, generate messages, suggest billing codes, or help with inbox work. That tool may make the physician more efficient. But it often has little contact with the patient's lived context: the symptom pattern across weeks, the fear behind nonadherence, the tradeoff the patient is unwilling to make, the wearable trend the patient has been watching, or the practical constraint that makes a textbook plan fail at home. > So both sides become AI-enabled, but the relationship does not become AI-native. The patient-AI loop stays with the patient. The physician-AI loop stays with the physician. The EHR remains a separate institutional artifact. Everyone may have more information, but there is still no common case model that both patient and physician can work from. That is not a triangle. It is three people dribbling in different gyms. > That is the missing layer. AI should not simply be a private assistant to each side of the encounter. In AI-native care, it becomes a third node in the clinical relationship: a common interpretive layer that absorbs the patient's lived context, the physician's clinical constraints, and the medical evidence into the same working model. This does not mean AI becomes the decision-maker. It also does not mean AI finds a bland compromise between patient preference and physician judgment. The point is not to split the difference. The point is to preserve both perspectives and make the actual constraint set visible. > The patient may be saying: "I am scared of this medication because I had side effects, I do not feel heard, and my daily life data does not match the plan." > The physician may be saying: "This risk is real, the consequences are high, and we cannot let an unverified hypothesis create false reassurance." The third node should help translate those into a common clinical object: what is known, what is uncertain, what matters to the patient, what matters medically, what is unsafe, what options exist, and what needs physician judgment. > That is the difference between adding AI to old silos and building AI-native care. The goal is not patient AI versus physician AI. The goal is a patient, a physician, and an AI-supported case model that lets both sides see the same court. Why the triangle offense works as a metaphor For readers who do not care about basketball X's and O's, the important idea is simple: the [triangle offense](https://www.basketballforcoaches.com/triangle-offense/) is not a single scripted play. It is a structure for decision-making. It is closer to a grammar than a recipe. At its core, the offense creates spacing between players so the defense cannot easily trap one person or clog every passing lane. Once the spacing is set, the player with the ball reads the defense and chooses the best available action. That decision then changes what everyone else does next. The offense can keep flowing until the right opening appears. The classic triangle has a three-person side of the floor and a two-person action on the opposite side. The ball often enters through a "trigger" player whose pass determines the next movement. The point is not randomness. It is structured adaptability: each player has a role, but the next move depends on what the defense gives up. That is why the metaphor works for AI-native care. The goal is not to script every interaction between patient, AI, and physician. The goal is to create enough structure that each participant knows their role, enough spacing that one participant does not collapse the whole encounter, and enough flexibility that the right actor can take the advantage when the situation calls for it. In other words, the point is not to make the clinic cosplay as the 1996 Bulls. The point is to stop everyone from standing in the same corner. Phil Jackson's teams made this famous because the ball did not stick. No single player was supposed to dominate every possession. The system worked when players understood where to stand, when to pass, when to cut, and when to take the shot. Patient, AI, physician AI-native care needs a similar triangle: patient, AI, and physician. The patient brings lived experience: symptoms, goals, fears, preferences, constraints, daily patterns, family context, and the reality of what happens between visits. The patient is the only person living the illness continuously. AI brings synthesis: organizing records, extracting timelines, finding patterns, drafting questions, explaining concepts, comparing possibilities, and turning messy information into something the patient and clinician can work with. The physician brings clinical judgment: diagnosis, examination, prioritization, risk assessment, medical responsibility, treatment decisions, safety boundaries, and the ability to know when something plausible is actually dangerous. If you want the 1990s Bulls version of the metaphor, the patient is Dennis Rodman: not because the patient is chaotic, but because the patient owns the raw possession battle. The patient gathers the loose balls of lived experience: symptoms, friction, logs, side effects, dietary reality, sleep, stress, and the details nobody else can see. AI is Scottie Pippen: the point forward. It sees the floor, connects the action, advances the ball, and turns scattered inputs into an assist. It does not need to be the final scorer to be essential. The physician is Michael Jordan: the closer. When the possession becomes high-consequence, ambiguous, or clinically risky, the physician has to take responsibility for the final shot: diagnosis, escalation, prescribing, and the judgment calls where plausibility is not enough. The point of the analogy is not hierarchy. Rodman, Pippen, and Jordan were valuable because they were different. The triangle worked because their differences fit into a structure. > At the center is what we might call compounding care: the clinical value created when the patient's lived context, the AI's synthesis, and the physician's judgment reinforce each other instead of competing for control. Situational ISO inside the triangle But the triangle does not mean every possession is evenly distributed. This is where the basketball metaphor becomes more useful. The triangle offense was not anti-isolation. It created spacing so the right player could attack the right advantage. Sometimes the ball moved around the structure. Sometimes the structure created an isolation. Compounding care works the same way. The goal is not to make the patient, AI, and physician do the same work at the same time. The goal is to recognize who has the advantage on this possession. Nobody needs a three-person committee to summarize a lab trend. Nobody needs a chatbot taking the final shot on chest pain. When AI gets the ISO When the problem is synthesis-heavy, AI can get the ISO. AI should isolate on the work it is structurally good at: organizing records, extracting timelines, comparing labs over time, scanning symptom logs, preparing visit briefs, and finding patterns that would be hard for any human to hold in working memory. When the patient gets the ISO When the problem depends on lived experience, the patient can get the ISO. The patient should isolate when the most important data lives outside the chart: symptoms, goals, fears, triggers, constraints, family context, tradeoffs, and what actually happens between visits. AI can help the patient become a better historian, but the patient is still the source of that lived signal. When the physician gets the ISO When the problem carries clinical risk, uncertainty, or edge-case complexity, the physician gets the ISO. The physician should isolate when plausibility is not enough: diagnosis under uncertainty, physical exam integration, medication risk, contraindications, escalation decisions, rare presentations, and deciding what not to chase. > The danger is hero-ball medicine, where one actor tries to dominate every possession. The better model is situational ISO inside a larger triangle: the ball moves until the right actor has the right shot. > And the triangle itself may not be the final form. It may be the transitional structure that teaches medicine how to move before care becomes more continuous, asynchronous, and motion-based. That is worth keeping open. As patients, physicians, diagnostics, and AI systems become more capable, the relationships between them may become more liquid. The important thing is not to freeze the metaphor too early. The important thing is to move beyond isolated AI silos toward a clinical relationship where context, synthesis, and judgment can keep circulating. The metaphor should move as the medicine moves. The promise and the chasm The point is not to put AI between the patient and physician as a replacement. The point is to put AI into the triangle as a third participant in the information flow. In the old model, the patient often arrived with a story and the physician had to reconstruct the case under time pressure. In the triangle model, the patient arrives with a structured brief. The AI has helped assemble the context. The physician can spend less time extracting the basics and more time doing what physicians are uniquely trained to do: interpret, examine, weigh risk, decide what matters, and protect the patient from false certainty. > That is the promise. The danger is that many physicians will experience this as an affront to authority. There is a double standard here medicine has not fully admitted: physicians are increasingly comfortable using AI to improve their own work, while treating patient use of the same class of tools as recklessness. That critique would land harder if physicians were consistently more AI-literate than patients, but many are using LLMs with a similarly thin understanding of hallucinations, provenance, and failure modes. For now, the patients doing this are unusual. They are founders, programmers, researchers, data-literate professionals, quantified-self enthusiasts, or people desperate enough to become expert in their own condition. That will not remain true. The spreadsheet people always arrive first. Then the interface gets easier. The tools are becoming easier. Patients will not need to know Python to upload labs. They will not need to understand vector databases to build a personal health timeline. They will not need to be AI experts to ask, "What patterns do you see in my symptoms over the last three months?" The interface will become natural language. The data pipes will improve. The analysis will become more available. As that happens, a chasm will open between two kinds of physicians. > One physician will say: "Do not bring me ChatGPT." > Another will say: "Show me what it found, show me the source data, and let's decide what is signal and what is noise." The second physician will not be less rigorous. They will be more rigorous, because they will insist on provenance, context, and clinical validation. > This is the posture medicine needs. > Not credulity. Not paternalistic dismissal. Disciplined partnership. What data did the model use? What did it infer versus what is directly documented? What are the strongest alternative explanations? What would change management? What could be dangerous if acted on prematurely? What needs confirmation with exam, labs, imaging, or specialist input? What is the patient's actual goal? What AI-native patients will want AI-native patients will not simply want a physician who tolerates AI. They will want a physician who can work with AI-shaped information. That means physicians will need to become comfortable reviewing patient-generated summaries without assuming they are worthless. They will need to distinguish a useful timeline from a hallucinated conclusion. They will need to know how to ask what model was used, what documents were included, and whether the patient can show the raw data behind the summary. They will also need to coach patients on how to use AI safely. A useful physician response might sound like: > "Bring me the summary, but also bring the original labs and notes. Ask the model to separate facts from hypotheses. Ask it to list what would make each hypothesis less likely. Do not start, stop, or change medications based on the model. Use it to prepare better questions, not to self-prescribe." That is not giving up authority. That is exercising authority in a world where patients have more tools. > The physician's authority should not depend on controlling access to information. It should depend on the quality of clinical judgment. What physicians can learn from these stories Sid's story shows the power of relentless information organization. Even if most patients will never build a thousand-page care document or pursue bespoke cancer therapies, the underlying lesson matters: complex care fails when information is fragmented. The future patient will expect their physician to understand that a medical record is not enough. They will want a living care model. Amy Deng's story shows the power of patient-generated longitudinal data. Many clinical mysteries are not solved by a single visit because the signal lives across time: symptoms, meals, sleep, medications, cycle timing, activity, stress, blood pressure, glucose, environment, and recovery patterns. AI is well suited to help structure that mess. The physician's job is to determine which patterns are clinically meaningful. Together, these stories point to a new baseline: > The patient is no longer just bringing symptoms. The patient may bring a dataset. > The physician is no longer just receiving a narrative. The physician may be asked to evaluate an AI-assisted analysis. > The visit is no longer just history-taking. It may become a review of competing hypotheses. That is a better use of physician skill, if the physician is ready for it. Collaborative skepticism The most important physician skill in this environment may be collaborative skepticism. Skepticism alone is easy. A physician can dismiss AI outputs because they are imperfect. That is often true, but not sufficient. Collaboration alone is also easy. A physician can nod along with whatever the patient brings. That may feel patient-centered, but it can become unsafe. > Collaborative skepticism is different. > It says: "I am willing to take your work seriously. I am also responsible for testing it clinically." That posture respects the patient without outsourcing medicine to the model. It invites patient agency without pretending every hypothesis is equally valid. It makes room for AI while preserving the physician's duty to protect from harm. This is not extra work forever. Done well, it can make visits more efficient. A patient who arrives with a coherent timeline, medication list, symptom log, and focused questions is easier to help than a patient forced to reconstruct months of complexity from memory. Medicine has enough mysteries. The medication list does not need to be one of them. > But only if the physician knows how to receive it. Invite the patient's AI-generated summary, but ask for source documents. Ask the patient what they believe and what they are worried about. Separate observations from interpretations. Identify what would change management. Look for dangerous misses before interesting zebras. Convert the patient's work into a prioritized plan. Explain why some hypotheses are worth pursuing and others are not. Give the patient a safe next step. AI does not remove the physician. It exposes the physician. > AI will expose which physicians are using authority as a shield and which are using expertise as a service. Patients with AI tools will still need physicians. They will need them badly. They will need someone to examine them, order the right tests, interpret uncertainty, understand base rates, manage risk, prescribe safely, coordinate care, and notice when the model is confidently wrong. Especially when the model is confidently wrong in complete sentences. But these patients will be less tolerant of unexplained dismissal. They will not accept "because I said so" when they have spent hours organizing their data and trying to understand their condition. They will want reasons. They will want partnership. They will want a physician who can say: > "That is an interesting pattern. Here is why I think it matters." > "That looks compelling, but I do not think it changes management because..." > "This part concerns me. We should not wait on that." This is the communication gap that will define the next decade of patient trust. > The best physicians will not compete with AI for authority. They will use AI to make the patient's story more visible, then bring judgment to it. The future visit Imagine a future visit where the patient comes in with three things: a clean timeline of symptoms, medications, exposures, diet, sleep, and relevant life events; a set of AI-generated hypotheses, each linked to supporting and contradicting evidence; and a short list of questions the patient wants answered. The physician opens the visit differently: > "I reviewed your summary. Let's start by checking the assumptions. Then I want to examine you, look at the raw labs, and decide what actually needs action." > That is the triangle working. > The patient is active. > The AI is useful. > The physician is essential. > The ball keeps moving. The mindset shift Physicians do not need to become AI maximalists. They do not need to endorse every patient-generated analysis. They do not need to pretend that LLMs are reliable clinicians. They should be very clear about the risks: hallucinations, overtesting, anxiety, false reassurance, inappropriate self-treatment, privacy, and inequitable access. But physicians do need to understand that the patient-AI relationship is becoming part of the clinical encounter. The question is whether physicians will meet it with irritation or leadership. The portal message is coming either way. The physician of the AI era should be able to say: > "Use AI to organize your story. Use it to learn. Use it to prepare questions. Use it to notice patterns. But bring me the evidence, not just the conclusion. My job is to help you determine what is real, what is risky, what matters, and what we should do next." > That is the triangle offense for AI-native care. > Not patient versus physician. > Not AI instead of physician. > Patient, AI, physician. > Spacing. Movement. Judgment. > And at the center: compounding care. Sources Century of Bio: [Sid](https://centuryofbio.com/p/sid) - Elliot Hershberg's essay on Sid Sijbrandij's patient-led oncology information stack. Metalearn: [I Solved My Mystery Fatigue with AI](https://metalearn.substack.com/p/i-solved-my-mystery-fatigue-with-ai) - Amy Deng's account of using AI for structured symptom and data analysis. Basketball For Coaches: [Triangle Offense Complete Coaching Guide](https://www.basketballforcoaches.com/triangle-offense/) - a clear primer on spacing, reads, trigger passes, and triangle principles. What is the triangle offense for AI-native medicine? It is a metaphor for care where patient context, AI synthesis, and physician judgment keep moving together instead of staying in separate patient-AI and physician-AI silos. Does this mean AI replaces the physician? No. AI can help organize information and surface patterns, but the physician remains responsible for examination, risk assessment, diagnosis, prescribing, escalation, and clinical judgment. What should physicians do when patients bring AI-generated summaries? Use collaborative skepticism: review the summary, ask for source data, separate facts from hypotheses, identify what would change management, and protect the patient from premature or unsafe conclusions.