# Now the Patient Has a Chatbot

*On the answer that arrives first, and the worry that was never typed*

Kanwar Partap Singh Gill, MD · November 5, 2026 · The Patient’s Side of the Screen

Picture a visit, imagined for this essay. A woman in her early sixties sits down, and before anyone has asked why she has come, she turns her phone around and holds it out. On the screen is a conversation she had with a chatbot a little after one in the morning. She had typed that her fingers go numb at night, often enough to wake her, and that she has to shake her hands before the feeling comes back. The answer is long, calm and well arranged. It lists several common explanations, asks a few sensible questions, advises her to see a clinician, and, near the end, in a single sentence, mentions rarer neurological causes.

She has read that sentence many times. You can tell from the way the screen has been scrolled to it.

The answer on the screen runs to several hundred words. The answer she has brought with her is one sentence long. That gap, between what the software said and what the person kept, is where this essay begins, because it is where the visit begins. Whatever the chatbot's answer was worth as information, it has already done something else: it has arrived first. It met her worry before anyone in a clinic did, gave it a vocabulary and a shape, and handed her a sentence to carry through the rest of the night. By the time she sits down, the conversation she came to have is no longer the first conversation about her hands. It is the second.

The hour matters too. Anyone who has lain awake with a symptom knows that the night enlarges things. By day a tingling hand is an inconvenience; at one in the morning, with the house quiet and nothing to do but attend to it, it becomes a question about the future. That is exactly the hour at which no clinic is open and the chatbot is. So the answer she received was not read in the conditions in which most medical information is meant to be read, with a person beside it and the day's ordinary proportions intact. It was read in the dark, by someone already afraid, and it was read the way frightened people read, which is selectively.

I want to think about both halves of that arrival. One is the information, which can be checked. The other is the relationship the tool appears to offer before any clinician enters the room, which cannot be checked in the same way and which may matter more to how the visit goes.

## The Oldest Second Opinion

Nobody has ever arrived at a doctor's door with an empty head. People have always come carrying explanations gathered in between: from books, from relatives, from a neighbour who once had something similar, and for a generation now from search engines. In Punjab, as across much of the subcontinent, the field of second opinions has long been crowded. There is the ਹਕੀਮ (*hakīm*), trained in the Unani tradition, and the ਵੈਦ (*vaid*), in the Ayurvedic. There is the compounder from the dispensary, who has learned a good deal by watching and is consulted accordingly. There is the ਸਿਆਣਾ (*siāṇā*), literally a wise one, who may offer a charm, a knotted thread or a remedy whose ingredients are not disclosed. And there is always the neighbour.

Physicians trained in the modern way have often been dismissive of all this, sometimes with reason. Confident authority without evidence is dangerous wherever it speaks from, a village courtyard or a hospital corridor. But the folk healer has properties worth naming, because the chatbot shares some of them and lacks others. The healer is available. He answers in the patient's own language, without forms. He does not ration his patience by the clock. And he lives in the village. He knows the family, his reputation is the sum of his results as remembered by people who can find him, and when he is wrong, someone knows where he lives.

Medicine has also long answered people it could not see. In the eighteenth century physicians of reputation, William Cullen in Edinburgh among them, carried on a large practice by post. A patient, or a local practitioner writing on the patient's behalf, set down the case in a letter, and the physician replied with an opinion and a regimen, having never laid eyes on the person described. Those letters are a reminder that a written self-description has always been a peculiar kind of evidence: a history in which the writer chose every word. The physicians who answered them knew it, and the careful ones asked for more, or leaned on the local practitioner who had actually examined the case.

The comparison exposes something about the chatbot that is easy to miss. We call it a second opinion, and in one sense it is. But a second opinion in medicine is valuable to the extent that it is independent: a different pair of eyes on the same patient, free of the first observer's framing. The chatbot cannot be independent in that way. It has not looked at her. It has looked at her account of herself, written at one in the morning, and everything it says is an elaboration of that account. It is less a second opinion than her own first opinion, returned to her organised, expanded and in a confident voice. That is not nothing; a person often understands her own worry better once it has been set out in order. But it is a different thing from being seen by someone else.

The Punjabi qissa understood something else about healers, which I find even more useful. In Waris Shah's *Heer*, the eighteenth-century telling of the lovers' story, Heer has been married against her will into the Khera household, and Ranjha has come to the village disguised as a jogi. To reach him she feigns a snakebite. The household's remedies and charmers fail, as they are meant to, and the jogi is summoned as a healer. The episode is about love and stratagem, not medicine, and I do not want to press it into a parable. But it contains a truth every clinician eventually learns: the complaint that brings someone to a healer and the reason they have come are not always the same thing. The healer who matters is the one who understands what the summons is for.

What the chatbot adds to this crowded field is a particular kind of company. It answers at one in the morning. It does not sigh, or glance at a clock, or have someone else waiting. It lets a person ask the same question three different ways until the answer seems to fit, and it never suggests that the question is foolish. In a 2024 poll by KFF, about [one in six American adults](https://www.kff.org/health-information-trust/poll-finding/kff-health-misinformation-tracking-poll-artificial-intelligence-and-health-information/) said they used AI chatbots at least once a month for health information and advice, rising to a quarter of those under thirty; yet most adults who had used or interacted with AI were not confident that chatbots' health information was accurate. That combination interests me. People are not consulting these tools because they trust them. They are consulting them because the tools are there, at the hour when nothing else is.

## A Program Called ELIZA

The appeal of a patient machine is older than the machines that now possess it. In 1966 Joseph Weizenbaum, a computer scientist at the Massachusetts Institute of Technology, described in the *Communications of the ACM* a program he called ELIZA, after the flower girl in Shaw's *Pygmalion* who is taught to speak like a lady. ELIZA ran on scripts, and the most famous of them, called DOCTOR, imitated a psychotherapist of the Rogerian school. Weizenbaum chose that role with some cunning. Such a therapist can reflect a patient's statements back without needing to know anything about the world; the reflection itself seems like understanding. In the transcript he published, the person at the keyboard tells the program that men are all alike, and that her boyfriend made her come here. The program answers, in the capital letters the paper used for its replies: "YOUR BOYFRIEND MADE YOU COME HERE". It had detected a keyword, transformed the pronouns and handed the sentence back. Nothing more.

What happened next troubled Weizenbaum for the rest of his life. In *Computer Power and Human Reason*, published in 1976, he described how people who knew perfectly well that ELIZA was a program confided in it anyway. His own secretary, who had watched him build it over many months, began a conversation with it and after a few exchanges asked him to leave the room. Some psychiatrists proposed that programs of this kind be developed into automated therapy that could reach many patients at once. Weizenbaum was appalled. He had not realised, he wrote, that brief exposure to a simple program could induce "powerful delusional thinking in quite normal people". The book argues that there are tasks a computer ought not to be given, whether or not it can be made to perform them, and it draws a distinction I have found helpful ever since: between deciding, which can be reduced to calculation, and choosing, which in his account requires judgment and, I would add, someone who will live with the result.

The modern chatbot inverts ELIZA's trick, and the inversion is instructive. ELIZA said less than you did. Its entire art was to return your words with a question mark attached. The systems people consult now say much more than they are told: they elaborate, organise, enumerate, qualify and advise. ELIZA was a mirror; the new tool is a library that talks. Yet the feeling the two produce is related. Weizenbaum's secretary felt heard by a machine that could only repeat her. The woman at one in the morning feels heard by one that seems to know everything.

There is evidence that the feeling travels into judgments about quality. In a widely discussed 2023 study in *JAMA Internal Medicine*, John Ayers and colleagues asked health professionals to compare physicians' and a chatbot's replies to questions posted on a public online forum. The evaluators [preferred the chatbot's replies](https://htn.co.uk/2023/05/02/researchers-evaluate-physician-and-chatbot-responses-to-patient-questions/) most of the time and rated them higher for empathy; the chatbot's answers were also several times longer. The authors were careful to say that they did not know how such answers would fare in a clinical setting, and I read the study as evidence about how answers read, not about what becomes of the people who receive them. Empathy judged from a transcript is a property of prose. It is still a property worth having, and the study tells me something uncomfortable about what people miss when a reply from a hurried physician is three lines long.

But companionship as it is felt and a therapeutic relationship are different things, and the difference is not sentimental. A relationship in the clinical sense has two parties, each of whom can be changed by it and each of whom can be held to it. The clinician who says *call me if it gets worse* has made a promise that someone can find him to keep. The chatbot has made no promise, and its memory of the conversation is whatever its makers have arranged. None of that makes the comfort it gave at one in the morning unreal. It means only that the comfort came from something that will not be there to answer for it.

## The Imitation Game

In 1950, sixteen years before ELIZA, Alan Turing published "Computing Machinery and Intelligence" in the journal *Mind*. It opens by proposing to consider the question "Can machines think?", and then sets the question aside as too ill-defined to be worth arguing about. In its place Turing put a game. An interrogator, communicating by teleprinter with two hidden players, tries to determine which is the man and which the woman; then a machine takes one of the parts, and the question becomes whether the interrogator can tell. Turing predicted that within about fifty years a machine would play so well that an average interrogator would have no more than a seventy percent chance of making the right identification after five minutes of questioning.

The sample exchanges he imagined are among the most charming passages in the history of science. Asked to write a sonnet on the Forth Bridge, his imagined respondent declines, explaining that it never could write poetry. Asked to add 34,957 and 70,764, it pauses about thirty seconds and gives an answer that is wrong by exactly one hundred. Readers still argue about whether Turing planted the error. Either way, the imitation he described includes hesitation and mistake, because a person would show both.

Turing's standard was conversational indistinguishability, and in short exchanges many of today's systems meet it handsomely. That is precisely why it is the wrong standard for medicine. A clinical answer is not judged by whether it sounds like a person. It is judged by whether it is right for this person, and by whether it knows what it does not know about her. Turing's imagined machine disguised its competence with a pause and, perhaps, a planted mistake. A modern system can disguise the opposite: its fluency is the same whether its content is sound or not, which is why federal guidance on generative AI counts confidently stated falsehood among the technology's characteristic risks.

Some of the published evidence about these systems is measured on examination questions modelled on medical licensing tests, and the scores can be impressive. I do not dismiss them; they show that a great deal of medical knowledge has been absorbed and can be retrieved. But an examination question arrives with the relevant facts already selected, in the right order, and with the distractors written by someone who knew the answer. A patient does not. She arrives with herself, in whatever order the night has left her, often with the most important part last.

If I were to propose a test of my own, it would not ask whether a system can pass for a person, or whether it can choose the right answer from a list. It would give the system a frightened person's first sentence and see whether it asks for the second. The best clinicians are distinguished less by the answers they know than by the questions they think to ask before answering, and a tool built to answer will, by its nature, be tempted to answer at once. Some systems are now designed to ask clarifying questions first, and that seems to me a step in the right direction. But a question asked by a program still depends on the person at the keyboard to recognise which of her details is the important one, and at one in the morning that is the very thing she is least able to judge.

## What the Prompt Left Out

Return to the imagined visit. Suppose the clinician, before reading the screen, asks a small question: what does *numb* mean, for you?

Here the conversation turns, as conversations do. It is not that she cannot feel her fingers, she says. They feel thick and clumsy, as though she were wearing gloves, and lately pins have been slipping out of them. She takes in alterations at home; her hands are part of how she earns her living. None of that was typed. The software answered, carefully and reasonably, the question it was given. In the room the question has shifted a little, and with it what an examination needs to look for and what matters most in the plan.

A prompt is a history taken by the patient, of herself, at one in the morning, without anyone to ask the second question. It is shaped by what she thought relevant, which is shaped in turn by what she feared and by the words she had. Some omissions are simple: the medicines she takes, the other conditions she lives with, how long this has been going on, whether one hand or both. Some are ambiguities of vocabulary, like *numb*, which means different things to different people and sometimes different things to the same person on different nights. And some are omissions of motive. She did not type that she sews for a living, because she did not think a computer would care, and she did not type what she was really afraid of, because typing it would have made it more real.

The omissions do not all run one way. People will sometimes tell a machine what they would not tell a person. In a 2014 experiment by Gale Lucas and colleagues, participants interviewed by an animated virtual figure disclosed more readily when they believed it was fully automated than when they believed a person was operating it. The finding fits ordinary experience. A program does not raise an eyebrow, does not know your neighbours, and will not be sitting across from you next year; Weizenbaum's secretary wanted the room to herself for a reason. So it is quite possible that, on some nights, the chatbot receives a more candid history than the clinic will: the drinking, the missed doses, the fear named outright, the symptom too embarrassing to describe face to face. That turns the usual worry on its head, and it suggests a question a clinician might learn to ask, gently and without any demand to see the screen: is there anything you told it that you haven't told me?

None of this is the chatbot's fault in any moral sense. Given those words, it did what a careful reader of words would do. The difficulty is structural. A system that answers what it is told depends entirely on what it is told, and the person telling is the person least able, in that hour, to know which details matter. A clinician starting from the same words would be in the same position. The difference is that a clinician can ask, look and touch, and is expected to.

## Four Ways to Be Wrong Politely

When people ask whether a chatbot's medical answers are accurate, they usually imagine a single score, like a percentage on a test. Accuracy in an answer of several hundred words is several different properties, and an answer can have some and lack others. Consider four kinds of failure, illustrated with invented fragments and not drawn from any real system.

The first is plain factual error: a sentence that states something false, such as *this test is not affected by any medicines*, when some medicines do affect it. This is the failure everyone expects, and it is in some ways the easiest to catch, because a fact can be checked against a source.

The second is unsupported certainty: *this is almost certainly a minor problem*. The sentence may even turn out to be true. But the information the system was given could not support *almost certainly*, and a reader has no way to see that the confidence was manufactured by the prose.

The third is misplaced emphasis. Every fact in the answer may be correct and the proportions still wrong: the common, benign possibility buried in a list, the rare and frightening one given a sentence of its own at the end, where a reader at one in the morning will find it. Emphasis is partly a property of the reader as well as the text. The printed order of an answer is not the emotional order in which it is read. Our imagined patient did not misread anything. She read the whole answer and kept the sentence that frightened her, which is what frightened people do.

The fourth is the gravest and the hardest to see: unsafe omission. An answer can be accurate in every sentence it contains and still leave out the circumstance in which waiting would be a mistake, the sign that should send someone for care tonight and not next week. An omission leaves no trace on the page. No reader can check a sentence that is not there.

This is one reason I distrust single accuracy figures, even honest ones. A study that reports a system's performance across hundreds of questions describes an average over questions. A person receives one answer, to her question, on her night. An answer that is right nine times in ten is a respectable instrument for a population and no comfort at all to the tenth reader, who has no way of knowing that she is the tenth.

What I take from this is that a coherent answer can fail in ways its coherence conceals. The answer on her phone was not wrong, as far as it went. It answered a word, and the word turned out to mean something slightly different to the person who used it. Fluency verifies nothing. Every medical claim on that screen still has to be weighed against her history, her examination and the evidence, by someone who will answer for the weighing.

## The Oracle's Benchmark

The ancient world had its own consultations at a distance, and one story about them seems to me almost indecently relevant. Herodotus tells, in the first book of his *Histories*, how Croesus, king of Lydia, considering whether to make war on Persia, decided first to test the oracles. He sent messengers to several of them with instructions to ask, on the hundredth day after setting out, what the king was doing at that moment. Croesus had chosen something no one could guess: he was boiling a tortoise and a lamb together in a bronze cauldron with a bronze lid. The oracle at Delphi answered correctly. Croesus, satisfied, sent lavish gifts and asked his real question. The answer came that if he attacked the Persians he would destroy a great empire. He attacked, and destroyed his own. Later, a captive, he sent his chains to Delphi in complaint. The reply, as Herodotus gives it, was that he ought to have sent again to ask which empire the god meant.

Croesus ran what we would now call a benchmark, and the oracle passed it. The high score did not transfer to the decision, because a test of whether a source can answer a question with a known answer is not a test of whether it will answer your question in a way you understand. And the fatal answer was not even false. It was correct, and it was misread, and the follow-up question that would have saved a kingdom was never asked.

Heraclitus, in a fragment that survives only because later writers quoted it, said that the lord whose oracle is at Delphi neither speaks out nor conceals, but gives a sign. That is the common English rendering, and I find it exact. The oracle's ambiguity at least announced itself. A questioner who received a riddle knew he was interpreting. The chatbot does the reverse: it speaks out, in complete and confident sentences, and it does not signal when it has answered a slightly different question from the one in the asker's mind. Its clarity can hide the interpretation that the oracle's obscurity forced into view.

Ancient writers report that visitors to Delphi passed maxims inscribed at the temple, of which the most famous is *know thyself*. It is usually read as a moral counsel, and it is one. It can also be read as a note on the limits of consultation. The god could answer a question; he could not supply the self-knowledge required to ask it well, or to recognise which empire was one's own. A chatbot is in the same position, and so, for that matter, is a physician. Neither can know what the person in front of them has not yet found words for. The difference is that a physician has a room, some minutes and a body to examine, and can help the words be found.

There is a further lesson in the oracle's machinery. The Pythia did not speak alone; there were priests, procedures, a sanctuary and centuries of practice behind her words, and ancient writers and modern scholars alike have argued about how much of any answer was hers. A chatbot's answer also belongs to a system. It depends on the particular model and its version, the instructions its operators have given it, whether it retrieves documents or answers from what it absorbed in training, what safeguards are layered on top, and the date on which the question was asked. The same question put twice may not receive the same answer. So a research finding about one system, tested in one way at one time, cannot be carried over to every public chatbot as though they were a single oracle with a single reputation. The study of forum replies measured how answers read to evaluators. The KFF poll measured use and trust, not accuracy. When the World Health Organization issued guidance on large multimodal models in January 2024, it listed patient-guided use, people investigating their own symptoms and treatment, among the applications it addressed, and it warned that such systems can produce false, inaccurate, biased or incomplete statements; its recommendations were addressed to governments, developers and health systems, and they are governance advice, not evidence of what happens to patients. Each piece of evidence answers its own question. None of them answers hers.

## Receiving the Phone

So how should the phone be received? I can imagine two unhelpful ways. One is to wave it off because of where the paragraph came from, which forces her to defend a piece of software when she came to talk about her hands. The other is to treat the paragraph's neatness as evidence that it understood her, and to let it set the agenda for the rest of the visit. Either way the software would be deciding too much about the conversation.

The most useful first question may not be whether the answer is correct. Imagine the exchange, still invented, going something like this.

*"I looked this up last night. I know doctors probably hate that."*

The apology is an invitation, and it is also a test: is this a place where I will be judged for having been frightened at one in the morning?

*"I'm glad you did. Which part of it stayed with you?"*

That question concedes no authority to the software. It locates the part of the conversation that now needs attention. A clinician could spend several minutes checking every line on the screen and still miss the reason the phone was offered. A person may bring a long answer because one sentence frightened her, because another gave her permission to seek help, or because she wants someone to explain why two of them seem to contradict each other.

*"This, at the end. About the nerves."*

*"What did you take it to mean?"*

Now the fear can be spoken, and the claim on the screen can be examined as a claim, against her history and her examination, with the reasons given aloud. There is a temptation, at this point, to correct the source before hearing the person. *These systems make mistakes* may be true and still arrive too early. It leaves the patient with an awkward choice: defend the tool that answered at night, or abandon the question she used it to ask. A better conversation examines a claim without turning its owner into the accused. When a correction is needed, it can be made without a loser: *The answer you got was a careful answer to the question you typed. Now that I've heard more, the question is a bit different, and here is what I think it means for you.* That sentence does not flatter the software, and it does not humiliate the person who consulted it. It treats her search as diligence, which is what it usually is.

And when the answer was good, as it often may be, the visit should say so. If the explanation on her screen was sound, telling her so does two things at once: it confirms what she has learned, and it shows her that the clinician's judgment is not reflexive opposition to the machine. Trust is easier to give to someone who can be seen agreeing as well as disagreeing. A clinician who treats every outside answer as an error teaches patients to stop bringing them, and the answers do not stop. They merely become invisible.

I should give the other side its full weight, because a clinician can underestimate a tool precisely because it is unfamiliar, just as a patient can overestimate it because it is attentive. Used with care, these systems may do real good around a visit. A person can use one beforehand to sort out what she wants to ask, so that the visit is spent on her real concerns. She can use one afterwards to have an explanation restated in plainer words, or in her own language, and then bring back whatever still does not make sense. She can ask it what a term in her note means. Some patients may already arrive better prepared than patients ever have, and that would be a gain, not a threat.

The limits are as practical as the uses. What is typed into a public chatbot may be stored and seen by others, a question another essay in this series takes up at length. Whatever such a tool says about a particular person's body needs checking with someone who knows that person. And it is not the place for a symptom that feels urgent to wait for an answer. Those are not technical caveats. They follow from the same distinction as everything else here: the words on the phone can be checked, but the circumstances that made those words matter have to be discovered, and discovering them is the visit's work.

## The Question Never Typed

In the visit I am imagining, the conversation reaches the thing she did not type. It was never really the nerves. It was the pins sliding out of her fingers, and what that might mean for the hems and seams that pay for her rent, and the thought, at one in the morning, of hands that could no longer do what they had done for forty years. The rare cause in that last sentence of the chatbot's answer now gets what it needed from the start: it is either addressed or put back in proportion, explicitly and out loud, so that it does not ride home with her. She leaves knowing what we think is happening, what we do not yet know, what will happen next and whom to call if something changes.

Something else happens in the visit that the night could not provide, and it is easy to overlook because it is so ordinary. Someone looks at her hands. Someone takes them, tests what they can feel and how strongly they grip, watches her fasten and unfasten something small, and says aloud what he finds. An examination gathers information, but it is also a form of attention that text cannot simulate, because it consists precisely of a person's body being taken seriously by another person in the same room. Whatever it shows, it answers a question the chatbot could not even receive: has anyone actually looked?

I suspect the shape of patients' questions will keep changing. More of them may arrive pre-sorted, already in the vocabulary of whatever answered them the night before: not *my hands go funny at night* but a technical term learned an hour earlier. That will often help. It also carries a small risk, because a borrowed term can bring a borrowed conclusion with it, and the original words, the ones that would have told the most, may never be spoken. Part of a clinician's work will be to ask for those original words back.

Croesus received an answer that was true and ruinous, and his failure was not that he consulted the oracle. It was that he did not go back and ask what it meant. The woman in this imagined visit did what he did not. She carried the answer to someone who could ask her the second question, and the visit's task was to ask it: not *is this right?* but *which empire did you think it meant?*

Somewhere in the middle of all this the phone goes back into her bag, face down. Neither of us remarks on it. Perhaps she will ask it something else tonight. I hope she does, and I hope that what she types next includes the part about the pins.

## Evidence and Editorial Note

**Documented findings.** The essay relies on the author's public biography: a family physician practising in California since 2010, who grew up in Amritsar and whose first language is Punjabi. About 17 percent of US adults using AI chatbots at least monthly for health information and advice (25 percent of those under 30), and a majority (56 percent) of adults who use or interact with AI not confident in the accuracy of chatbot health information, from a survey conducted June 3–24, 2024: KFF Health Misinformation Tracking Poll (linked). Health-professional evaluators preferring chatbot replies to physician replies in 78.6 percent of 585 evaluations of 195 questions from a public online forum, rating them higher for quality and empathy, with chatbot replies considerably longer, and the authors' caution about clinical settings: Ayers and colleagues, *JAMA Internal Medicine*, 2023, as summarised by HTN (linked). Confidently stated false content as a recognised risk of generative AI: National Institute of Standards and Technology, Generative AI Profile (NIST AI 600-1), July 2024, mentioned only in passing. The World Health Organization's guidance on large multimodal models (January 2024), its inclusion of patient-guided use and its warning about false, inaccurate, biased or incomplete statements: from the published guidance as widely reported; Joseph Weizenbaum, "ELIZA—A Computer Program for the Study of Natural Language Communication Between Man and Machine", *Communications of the ACM*, January 1966, its DOCTOR script and published transcript (one short quotation); Weizenbaum, *Computer Power and Human Reason* (1976), the secretary anecdote, the proposals for automated therapy, the deciding/choosing distinction, and one short quotation. Alan Turing, "Computing Machinery and Intelligence", *Mind*, 1950: the opening question, the imitation game, the fifty-year and seventy-percent prediction, and the sample sonnet and addition exchanges, described. Herodotus, *Histories*, Book 1, on Croesus's test of the oracles, the "great empire" response and the later reply about which empire was meant; Heraclitus's fragment on the Delphic oracle, in a common English rendering. Waris Shah's *Heer*, the feigned snakebite and the summoning of the disguised Ranjha, described from the traditional text. William Cullen's consultation practice by letter in eighteenth-century Edinburgh, described in general terms from the history of medicine. Gale Lucas and colleagues, "It's only a computer: Virtual humans increase willingness to disclose", *Computers in Human Behavior*, 2014, described without figures. Ancient testimony to the maxims inscribed at Delphi, including "know thyself". The figures of the hakim, vaid, compounder and siāṇā are described as general features of healing in Punjab, not as claims about any person. No error rate, clinical benefit or harm is asserted for any tool. **Vendor representations.** None; no chatbot, model or company is named.

**Recollection and tradition.** The visit, the patient, her symptom, her work, her conversation with a chatbot and the dialogue in the clinic are hypothetical and illustrative; they do not describe a real patient or a real conversation. No particular healer, consultation or family occasion from the author's life is described. No change in real patients' phrasing is reported from the author's practice; the remark about pre-sorted questions is speculation.

**Analysis and forecast.** The reading of chatbot use as a relationship question, the readings of ELIZA, Turing, Herodotus, Heraclitus and Waris Shah as they bear on it, the four kinds of failure, the suggested wording of a correction and the forecast about how questions may change are the author's personal interpretation and preference as of November 6, 2026. The essay describes no diagnostic or treatment approach. **Independence.** This essay is not sponsored by any institution, company, health system or technology developer. It is personal reflection, not medical advice; readers seeking care should contact their own physician's office.



## Related reading

- [The Art of Saying “I Don’t Know Yet”](https://personal.kpsgill.com/essays/the-art-of-saying-i-don-t-know-yet/)
- [When Patients Started Reading Their Notes](https://personal.kpsgill.com/essays/when-patients-started-reading-their-notes/)
- [Wrong Beautifully](https://personal.kpsgill.com/essays/wrong-beautifully/)