Paternalism
in Medical AI

The title performs the entire premise. A patient question, already answered by the system the editorial is critiquing, is escorted from the consultation room, handed to a physician for translation, and returned to the patient. The patient remains where they were before they asked.

The editorial does not address why the patient was asking ChatGPT instead of the physician. The wait is invisible to the argument. It exists outside the frame in which the argument can be made. The reader is invited to imagine a counterfactual world in which the physician was available, in which the visit was long enough, in which the question could have been asked the way the editorial recommends.

The counterfactual world does not exist for most of the patients who are using ChatGPT. The editorial is written for a clinical reality the patient could not access. That is the artifact.

The verb "forcing" does work the reader is asked not to inspect. Nobody is forcing the chatbot to perform anything. Patients are asking it questions. The framing relocates the agency from the patient who asked to the technology that responded.

The list of threats is also worth reading carefully. Regulatory, risk management, professional. Each of these is an institutional category. The patient is not on the list. The threats inventoried are the threats to the people writing editorials about the threats. The threat to the patient who has no other way to be heard is not enumerated because it is not a threat to the editorial's primary audience.

"Authorized" is the operative word. The editorial is correct that nobody authorized ChatGPT to answer patient questions. The patient also did not authorize the institution to be the gatekeeper of those answers. The asymmetry of who needs whose authorization is the structure of the gallery.

The imperative falls on the patient. The robust understanding required, in the time available, by the means available, with the literacy available. The conditional governing all of this, what to do when a robust understanding cannot be obtained because the appointment is in November and the pain is now, is not addressed.

The literature on the median primary care wait time, the median visit length, and the median patient understanding of discharge instructions exists. The figures are not obscure. None of it appears in this conclusion. The frame contains the patient and the chatbot. The system that produced the need for the chatbot is offstage.

The methodological assumption is that the patient is asking the chatbot because the chatbot is convenient. The empirical pattern is that the patient is asking the chatbot because the alternative is silence. Until the literature accommodates the second possibility, the imperatives it issues will continue to be addressed to a patient who does not exist.

The Latin is doing the entire argument. Caveat emptor: let the buyer beware. The risk is on the patient. The mistake is the patient's. The harm, when it lands, lands on the patient who chose poorly. Then, carpe diem: build the future together. The future is on the institution. The collaboration is on the institution's timeline.

The split between the two phrases is the artifact. The patient bears the risk in the present tense. The institution bears the responsibility in the future tense. The future is undated. The risk is now.

The post is courteous, learned, allyship-coded. The author is on the patient's side, as the author understands sides. But the Latin admits what the courtesy does not. The system is asking the patient to absorb the cost of the system's unfinished work.

The argument is a good argument. The three capacities are real. The omission is the artifact. The patient also sits with uncertainty. The patient also carries responsibility, often in literal sole custody of their own outcome, with no malpractice insurance, no peer review, no senior colleague to correct a confident error. The patient also exercises judgment in the presence of real human stakes, where the stakes are their own life.

The framework is built for the clinician's encounter with AI. The patient is not in the framework. The conversation about composite intelligence, about humans contributing depth and judgment while machines contribute speed and breadth, ends at the consultation room door. On the other side of the door, the patient is alone with the output and is, by the framework's own terms, the least equipped person in the chain to interrogate it.

The exhibit is here not because the framework is wrong but because the framework's audience is incomplete. Extend it one person further down the chain and the entire premise of patient-facing AI safety changes.

The artifact in this case is not the institutional position but its counterweight. The patient describing the gap is the same patient the editorials in this gallery are addressing. They are educated. They are articulate. They are, by their own admission, asserting themselves. The institution still cannot meet them in time.

The pivotal line is the third sentence. Note that it is not a complaint about the provider's competence or care. It is a statement of fact about availability. The patient is not arguing that the LLM is better than the doctor. The patient is reporting that the LLM is present.

Read this exhibit alongside Exhibits 1.1 through 1.5. The editorials issue imperatives to patients about how to wait correctly. The patient issues a description of what happens when there is nothing left to wait for. The two literatures do not address each other. They occupy adjacent rooms in the same building and pretend the wall between them does not exist.

A chatbot saying that ibuprofen is an anti-inflammatory may or may not be in violation. Saying take 400 milligrams almost certainly is. The line between information and advice is left for litigation to settle, with the bill's private right of action providing the funding.

The legislative theory is that an AI system is performing the unauthorized practice of medicine when it gives advice a licensed person would have to be licensed to give. The theory does not require the AI system to be wrong. Accuracy is not the trigger. The trigger is the absence of a license.

What the bill is actually defending is the boundary of professional jurisdiction. The patient harm framing is the public-facing argument. The structural function is gatekeeping. Note the corollary: a chatbot that gives correct, well-sourced advice that happens to coincide with what a licensed professional would say is, by this bill's logic, in greater violation than a chatbot that gives vague non-answers.

The chief executive's email is the artifact, but the second-order artifact is what the email does not say. The directive bans clinician use because of confidentiality concerns. It does not address what should be substituted in the workflows the clinicians had been augmenting. The work returns to where it was. The reason the clinicians had reached for the tool returns with it.

This is the clinician-facing version of the same pattern that produces patient-facing bans. A technology is found to be in use without authorization. The authorized response is to forbid further use. The unauthorized use, which had been a workaround for something the institution had not solved, is treated as the problem rather than as the symptom of the problem.

One AMA representative described the order as having been triggered by "a lone medical geek." The phrasing is illuminating. The shadow user is described as deviant. The institutional inability to provide a sanctioned tool of comparable function is described as the baseline.

Read the two clauses together. The operator must disclose. The disclosure does not protect the operator. The user, having been told that they are talking to a chatbot, is judged unable to act on that information.

The doctrine being constructed is that the user's informed consent is not a defense against the user's harm. This is a coherent doctrine in cases where consent is presumed manipulated. It is a striking doctrine when the manipulation alleged is the user choosing to ask. A New Yorker who reads, in plain English, the disclosure that they are talking to an AI, is, by the bill's logic, still operating without informed consent. The remedy proposed is not to give the user better tools to act on the disclosure. The remedy is to remove the chatbot.

The corollary is also worth seeing plainly. A pharmacist who fails to give a patient the medication leaflet is in violation. A pharmacist who gives the leaflet is in compliance. The leaflet is the safe harbor. In the S7263 framework, the equivalent disclosure does no such work. The user's reading comprehension is not extended the same trust the user is extended at every other point in the consumer protection literature.

The asymmetry is the artifact. When the clinician uses the unsanctioned tool, the institution builds workflows around the behavior. When the patient does, the institution builds notices around the behavior.

The institutional theory is consistent. The clinician's unauthorized use indicates an unmet workflow need that the institution must solve. The patient's unauthorized use indicates a behavior problem that the patient must solve. The same behavior, performed by two different actors inside the same institution, generates two opposite responses. The variable is not the behavior. The variable is the actor.

The notice telling patients not to bring AI-generated information to the visit has no enforcement mechanism, which is part of its design. It does not need one. The patient is already inside an institution that controls the rest of the encounter. The notice is the polite form of the ban.

The exhibit is included for the through-line. The vocabulary is fifteen years older than S7263 and the rhetorical structure is identical. The patient is doing something they should not be doing. The doing produces stress. The remedy is to stop. The institution offers, in lieu of the prohibited resource, the institution itself.

The professional-training argument is the most durable. Ten years. The numbers vary by jurisdiction but the move is consistent. Whatever the patient is using is set against the full clinical curriculum and found inadequate. The comparison flatters the institution. It also flattens the question, because the patient is not in fact attempting to perform the function of a ten-year curriculum. The patient is attempting to ask a question.

The Australian doctors did not, in 2011, stop the patients from using Google. The patients used Google more. By 2020, 69 percent of Canadians used the internet for health information. The exhibit is here as a control case. A confident institutional ban, issued early, against a more limited technology, produced no measurable change in patient behavior. The exhibit predicts the half-life of its 2026 equivalents.

The exhibit is here because it is the structural ancestor of the present moment. A consumer-facing service that gave individuals direct access to information about their own bodies was ordered to stop. The harm cited was the risk that the consumer, on receiving the information, would misunderstand or misuse it. The proper place for the information was understood to be a clinician's office.

The legal theory was that the service was an unauthorized medical device. The harm theory was that the consumer was not equipped to receive the output without professional mediation. Read the two together. The same information, delivered through the same molecular test, was acceptable when a clinician ordered it and unacceptable when the consumer ordered it directly. The variable was not the information. The variable was the chain of authorization.

The intervening years have rendered the original concern empirically testable. The catastrophe of patients self-treating on the basis of 23andMe results did not occur at the scale predicted. What occurred was that millions of people learned things about their own genomes. The Bloom Syndrome carrier test was eventually authorized. The structure that produced the ban remains in place and now governs the regulation of patient-facing AI.

The skill required to operate the substitute is held against the patient. The skill required to operate the original is not.

The patient who phrases their symptoms poorly to a physician is given the visit anyway. The physician interprets, prompts back, examines, asks again. The patient who phrases their question poorly to a chatbot is, by this editorial's logic, disqualified. The asymmetry is invisible to the author because the institution is invisible to itself. The institution's own scaffolding around imprecise patient communication is naturalized to the point that it does not appear to the writer as a feature of the comparison.

The bracketed parenthetical is the small honest moment. The skill can improve. Which is to say, the skill is contingent, not constitutional. The patient who has not yet learned to prompt the chatbot is in the same position of learning as the medical student who has not yet learned to take a history. The editorial does not consider this analogy because the editorial does not consider the patient as a learner. The patient is rendered as a fixed quantity, evaluated against a moving target.

"Tendencies" is the operative word. Reading about your own condition becomes a tendency, a predisposition, a symptom. The lexicon of clinical pathology arrives to describe a patient using a search engine.

The verb "facilitates" carries the rest of the sentence. The AI is not informing the patient. It is enabling them. The frame inverts when the same access is held by the clinician, in which case the AI is "supporting decisions" and "augmenting clinical judgment." Same outputs, different verbs. The exhibit demonstrates how vocabulary is doing the work the evidence does not.

"Self-diagnosis" itself is worth unpacking. A patient who consults a chatbot and forms a hypothesis about their condition is, in the editorial's vocabulary, engaging in self-diagnosis. A patient who consults a clinician and is given a hypothesis is engaging in standard care. In both cases a hypothesis arrives in the patient's mind. The provenance is the artifact. The patient's hypothesis is treated as suspect by virtue of its origin. The clinician's hypothesis is not.

The post is among the most circulated of the genre. Read the first paragraph carefully. Three of the physician's own five diagnoses matched the chatbot's. The chatbot, in other words, was right on the differential. It was the patient's fear that the physician treats as the problem.

The diagnosis the post is making is on the patient, not on the chatbot. The chatbot performed within the standard of care. The patient performed outside the standard of composure. The implicit prescription is to remove the chatbot rather than to address the system that left the patient alone with it for two days.

The genre is large and continues to grow. The pattern across instances is consistent. The chatbot is accurate. The patient is upset. The upset is the problem. What is rarely asked in the genre is why the patient was alone for two days with a differential that included a life-threatening diagnosis. The institutional gap that produced the alone time is not the post's subject. The patient's reaction to being alone is.

The word is the artifact. The medical profession invented a clinical term to describe what patients do when they look up their symptoms. The term has the suffix of a syndrome. It enters the patient's chart as if it were a diagnosis.

Consider the inverse coinage that does not exist. There is no clinical term for the physician who orders a panel of tests because they cannot rule out a rare condition. There is no syndrome name for the specialist who runs imaging out of caution. The same information-seeking behavior, anxiety in the face of uncertain medical information, is virtuous diligence in the clinician and a pathology in the patient. The pathology has a name.

This is the cleanest exhibit in the gallery because the medicalization of patient information-seeking is built into a single word. Once "cyberchondria" exists, the patient who consults the internet is no longer engaged in a behavior. They are engaged in a condition. Conditions require treatment. The treatment, in this case, is to stop consulting the internet.

The line is doing two things. It is selling the credential. It is also redirecting the patient to the institution that holds the credential. The artifact is candid enough to make the structure visible. The blog is patient education, by which is meant patient direction. The direction is toward the system that produced the blog.

The implicit argument is that the patient should consult only sources that have been to medical school. The implicit population of sources that meet that bar is limited to physicians, who are the staff of the institution publishing the blog. The patient is being instructed that the only acceptable input is the input that flows through the institution.

The line is also doing something the institution does not say. The credential is being deployed to refute a tool that, when measured, frequently produces information that overlaps substantially with what the credentialed person would say. The credential is being asked to do work the comparative evidence does not support. The exhibit is here because the rhetoric is precise about its premise and inattentive to its evidence.

The artifact in this case is the word "anecdotal." It functioned as a dismissal in the 1990s. It functions, in different costume, in the 2026 editorials in this museum. What patients report from their own bodies and their own experience is, by the lexical choice, set against the structured data of the trial and found wanting.

The empirical record now is unambiguous. The patient communities that were called anecdotal in the 1990s identified the drug withdrawals, the missed diagnoses, and the underused treatments that the structured literature later confirmed. The vocabulary that flattened their findings was wrong. The vocabulary is still in service.

The exhibit is paired with the present-day editorial language for a reason. The continuity is the artifact. A profession that was wrong about patient knowledge in the 1990s and was wrong about it in the 2010s is now being wrong about it in the 2020s with greater conviction and a fuller vocabulary. The conviction does not prove the rightness. The conviction is the warning sign.

The rule produces a perverse incentive. A tool that gives the patient three options to choose between is, regulatorily, a calculator. A tool that gives the patient the right answer is a medical device. The cost of being correct is the years and millions required to enter the market.

The tools that survive in the patient-facing market are the ones that hedge. The hedge is encoded by the regulator and then read by the public as the system's honesty rather than as its scar. A patient receiving three options is told, by the structure of the offering, that the matter is uncertain even when the matter is settled. The uncertainty is regulatory, not clinical, but it arrives at the patient as if it were clinical.

This is a particularly subtle exhibit because it does not involve a ban. The regulator did not forbid the better tool. The regulator only made the better tool prohibitively expensive to bring to market. The patient receives the cheaper tool and is invited to act on its hedged output. The act of hedging is encoded into the patient's permission set without the patient being told that the hedge is a regulatory artifact.

The exemption is not about risk. The exemption is about who is in the room. The same software, with the same outputs, with the same accuracy, is exempted when a clinician reviews it and regulated when a patient sees it.

The doctrine being expressed is that the patient is the contaminant. Place a clinician between the patient and the tool and the tool becomes safe. Remove the clinician and the tool, unchanged, becomes dangerous. The danger is not in the tool. The danger is in the unmediated patient.

The structural consequence is one of the most consequential in patient-facing health technology. A startup building a tool for clinicians can move quickly. The same startup building the same tool for patients faces years of premarket review. The market for direct-to-patient tools is, by regulatory design, the harder market to enter. The tools that make it through are disproportionately those backed by capital sufficient to absorb the regulatory cost. Patient access to AI ends up controlled, in practice, by the funding environments of the firms able to afford the regulatory burden.

The label is doing the protection. The tool is more accurate than most clinicians on the task it performs. The patient who reads the result correctly is left with a clinically actionable finding the regulator declines to call a finding.

The structure is honest, in its way. The institution does not trust the patient to act on accurate information. It also does not trust itself to issue accurate information without a clinician between the device and the user. The label is the membrane.

The label has a downstream effect that is often missed. Because the device's most reliable output is described as not-a-diagnosis, the patient who shows up at the clinic with the result is in a doubly disadvantaged position. The result is real but the institution has authorized them to disregard it. The clinician, looking at the same result with the same numbers, can reach the diagnosis the device could not. The chain of authorization, which was the point of the label, is also the source of the patient's powerlessness over their own data.

The juxtaposition is the artifact. The agency feared the consumer would misunderstand and self-treat. The actual evidence of what patients did with the results was the opposite. They got another test. They talked to their physicians. They did, in other words, exactly what the regulator was demanding they be prevented from doing by accessing the original test.

This is the cleanest exhibit in the wing because it includes its own empirical refutation in the same paragraph. The harm anticipated did not materialize. The behavior anticipated did not materialize. The exhibit was not retired. The framework that produced the exhibit remains in place. New patient-facing technologies are routinely measured against the original fear rather than against the empirical record of what happened when patients were given the original tool.

The agency's prediction was a specific empirical claim about patient behavior. Patient behavior was the test. Patient behavior contradicted the claim. The framework that issued the claim treats the contradiction as a contingent fact rather than as evidence against the framework. The exhibit is here because regulatory humility is rare and this exhibit demonstrates the cost of its rarity.

The critique is, in some respects, valid. Sourceless output is a real problem of verification and patients should have access to citations. The exhibit is here because of what the critique leaves unsaid.

The standard the chatbot is being measured against is not, in fact, the standard of the consultation it is being told to defer to. Most clinical encounters do not produce a written list of citations for the patient. The clinician issues a verdict. The patient is expected to act on it. The citations are in the literature the clinician has consulted and in the training the clinician has received, both invisible to the patient. The patient's actual position in a consultation is also one of receiving confidence without sources.

The editorial does not propose that consultations should also be required to produce written citations. It proposes that the chatbot's sourceless output should produce greater patient deference to the institution. The institution gets to be sourceless. The chatbot's sourcelessness disqualifies it. The asymmetric standard of proof is the artifact.

The exhibit is included because it is a controlled experiment whose result is now in. Before clinicians were required to share their notes with patients, the predicted harms were inventoried. Read the list. Fear, guilt, anger, depression, confusion, frustration, hopelessness. The list is exhaustive.

The list is also, with respect to the empirical record, wrong. When the OpenNotes pilot ran, the predicted harms did not appear at scale. Patients reported feeling more in control. Patients reported better adherence. Patients reported improved understanding. Across three large health systems, at the end of the pilot year, no participating clinician opted out. The fear, in retrospect, was a fear about patient capacity. The fear was wrong about patient capacity.

The exhibit is paired with the patient-AI editorials in this museum because the predicted-harm structure is identical. A new technology gives the patient access to information previously held by the clinician. The clinical literature inventories the harms expected to follow. The empirical record then arrives and contradicts the inventory. The inventory is not retired. The pattern repeats with the next technology. The exhibit is the warning that the editorials in this museum are about to be wrong in the same way the OpenNotes warnings were wrong, and that the fact of the wrongness will not, by itself, be sufficient to retire the framework.