A psychiatrist on how artificial intelligence changed the first five minutes of every appointment, and why that might be a good thing.
There’s a question I ask near the beginning of almost every psychiatric evaluation.
What have you researched?
It’s a simple question, and for most of my career it produced a fairly predictable range of answers. Some people had read a Wikipedia article. Some had gone down a WebMD rabbit hole at two in the morning and arrived convinced they had something rare and terrible. Many had done nothing at all, which was its own kind of information.
Over the past year or so, the answer has changed. Not gradually. It feels closer to a phase transition. Now, when I ask that question, the response is almost always some version of the same thing.
I asked ChatGPT.
Sometimes it’s Claude. Sometimes Gemini. The specific tool matters less than what it represents: my patients are no longer arriving having read about their symptoms. They’re arriving having had a long, iterative, surprisingly sophisticated conversation about them, often over several nights, sometimes over months. They come in with vocabulary. With hypotheses. With medication questions I would not have expected from a first-time patient a few years ago.
They come in more informed. What’s striking is that they don’t always come in more certain.
The starting point moved
For most of the history of medicine, the physician was the primary source of medical information, and the appointment was where information transfer happened. You came in with a problem. You left with an explanation.
The internet complicated that, but not as much as people predicted. Search results are fragmentary and contradictory. Reading ten articles about fatigue doesn’t produce a coherent understanding of fatigue. It produces ten disconnected impressions, several of them alarming. Most patients who Googled their symptoms arrived not with knowledge but with anxiety wearing knowledge’s clothing.
Conversational AI is different in a way that matters clinically. It’s not a list of links; it’s a dialogue. Patients describe their symptoms in their own words, get a response, ask a follow-up, get another. They can say no, it’s not quite like that, it’s more like this, and be understood. They can ask the embarrassing question they’d never ask a person. They can ask the same question five different ways at three in the morning without anyone sighing.
What emerges from that process is something genuinely new: a patient who arrives with a model of their illness. Not just a symptom list, but a working theory about what’s happening to them and why.
That changes my job. I am, less often now, the person providing the information. I am, more often, the person helping someone interpret information they already have, validating some of it, refining some of it, and occasionally, carefully, correcting it.
Why I ask about AI on purpose
I want to be clear that asking about this isn’t a defensive maneuver. I’m not screening for misinformation so I can clear it away and get to the real work.
Asking what someone has researched is real work, arguably some of the most valuable minutes of a psychiatric evaluation.
Here’s why. In psychiatry, the presenting complaint is rarely the whole story. Someone says “I think I have ADHD,” and buried inside that sentence is a great deal of clinical information that has nothing to do with whether they actually have ADHD.
Why ADHD, specifically? What made that framework feel right? What were they hoping it would explain? What did they read that frightened them? What did they read that relieved them? What did they not ask the AI about, and why not?
When a patient tells me they spent two weeks reading about bipolar II, I learn something about their fear. When they tell me they researched every possible side effect of an SSRI before their appointment, I learn something about how they’ll approach treatment. When someone has clearly been searching for an explanation that would mean the problem isn’t their fault, I’ve learned something important about their self-narrative, and about what I’ll need to address before any medication has a chance to work.
Psychiatry runs on understanding a person’s mental model of their own suffering. AI conversations have become, for many patients, the place where that model gets built. Not asking about it means missing where the story actually started.
What AI is genuinely good at
I want to give AI its due, because I think dismissiveness here is both inaccurate and counterproductive.
Used well, these tools help patients do several things that make care better.
They organize. Depression and anxiety scramble memory and timeline. A patient who has spent an hour working with an AI to construct a coherent chronology (when the symptoms started, what changed, what medications were tried and for how long) arrives with something enormously useful. I’ve had patients bring in medication histories more accurate than the ones in their records.
They translate. Medicine has a vocabulary problem. Patients who have used AI to understand what anhedonia means, or what executive function refers to, or what an SSRI actually does at the receptor level, can participate in a different quality of conversation.
They prepare. Some of the best questions I get asked come from patients who worked out what they wanted to know before they arrived. Appointment time is finite. Preparation makes it go further.
They reduce shame. This one is underrated. There are questions people find genuinely humiliating to ask a human being: about intrusive thoughts, about sexual side effects, about substance use, about whether what they’re experiencing means they’re “crazy.” Getting a calm, non-judgmental first answer from a machine sometimes gives people the courage to raise it with a person.
They enable shared decision-making. A patient who understands the tradeoffs between two medications can actually participate in choosing between them. That’s not a threat to my authority. That’s the goal.
What AI cannot do
And now the other half, which matters just as much.
It doesn’t know your history. Not really. It knows what you’ve told it in a conversation, filtered through your own perception, missing the things you didn’t think to mention and the things you don’t know about yourself. It doesn’t have your labs, your prior records, your family’s account of what you were like six months ago.
It cannot examine you. It can’t observe your psychomotor activity, hear the flattening in your voice, notice that you’ve lost fifteen pounds, or register the thing you’re carefully not saying.
It confabulates. Language models generate fluent, confident text, and fluency is not accuracy. They can produce citations that don’t exist and drug interactions that aren’t real. The confidence is uniform whether the content is correct or invented, which is precisely what makes it dangerous in medicine. One Harvard study found that ChatGPT’s cancer treatment recommendations fully aligned with national guidelines in only 62 percent of cases, and in 12.5 percent produced recommendations entirely absent from the guidelines, including curative therapies for non-curable cancers.
It has no stake in the outcome. It won’t follow you for two years. It won’t notice that you’ve been subtly declining across three visits. It won’t sit with the discomfort of an uncertain diagnosis and keep thinking about your case in the car.
It cannot hold nuance the way clinical judgment does. Psychiatric diagnosis is not pattern matching against a criteria list. It’s weighing a life against a framework, repeatedly, with attention to what doesn’t fit.
The story everyone is sharing, and what it actually shows
Recently OpenAI publicized an account of a woman whose brain tumor diagnosis was revisited after her husband used ChatGPT to help him understand her pathology and imaging reports. The company described the AI as helping catch a misclassification, with the diagnosis revised from glioblastoma to IDH-mutant astrocytoma, a change with substantially different prognostic implications.
It’s a compelling story. It’s also more complicated than the version that circulated.
A physician involved in the case publicly disputed the framing, saying plainly that nobody had missed anything, and that the original 2020 diagnosis was correct under the tumor classification system in use at the time. The World Health Organization revised its central nervous system tumor classification in 2021. Under the newer framework, IDH-mutant tumors are recognized as behaving very differently from IDH-wildtype glioblastoma: growing more slowly, occurring in younger patients, responding better to treatment, and carrying significantly longer survival. The taxonomy changed. The pathology didn’t.
I find the corrected version more instructive than the marketing version, and I’d rather my patients hear it.
Because what actually happened is still remarkable. It’s just not “AI caught what doctors missed.” What happened is that a frightened family used a tool to understand a document they couldn’t otherwise read, noticed a molecular finding whose significance had shifted, and brought a question back to their oncology team. Physicians then reviewed the case and made a clinical determination.
That is precisely the right use of these tools. Not diagnosis. Not second-guessing. Better questions, brought to the people qualified to answer them.
Why psychiatry is a special case
Every specialty is dealing with this. I’d argue psychiatry is dealing with something slightly different.
In most of medicine, the objective data eventually adjudicates. The imaging shows what it shows. The culture grows what it grows. The patient’s interpretation matters, but it isn’t the diagnostic substrate.
In psychiatry, the conversation is the diagnostic instrument. There’s no blood test for depression. The evaluation is built almost entirely from what a person tells me about their emotions, thoughts, behaviors, timeline, relationships, stressors, and their own sense of what’s happening to them.
Which means that when a patient has spent forty hours in conversation with an AI about their inner life, that experience isn’t outside the clinical picture. It’s part of it. Their self-understanding has already been shaped by it before they walk in. Sometimes helpfully. Many patients have genuinely clarified their own history through that process. Sometimes less so. I’ve seen people arrive with a diagnosis they’ve become attached to in a way that makes it harder to see the actual pattern.
Either way, I need to know. Not knowing means evaluating someone’s mental model while pretending it appeared from nowhere.
What good care looks like now
The most common thing I encounter is some version of: I think I have ADHD. Or bipolar disorder. Or PTSD. Or: ChatGPT said this medication could be causing my symptoms.
There are two bad responses to this.
The first is reflexive dismissal: the eye-roll, the well, let’s let me be the doctor here. This is condescending, and it’s also clinically wasteful. Patients are frequently right. Adult ADHD is genuinely underdiagnosed. Bipolar spectrum illness is genuinely missed for years. Medications genuinely cause the side effects patients suspect. A physician who dismisses patient hypotheses on principle will miss things.
The second bad response is reflexive acceptance: writing the prescription because the patient arrived certain. This isn’t respect; it’s abdication. Anxiety, depression, ADHD, trauma, sleep disorders, thyroid disease, and substance use overlap enormously in presentation. Certainty is not evidence.
The right response is neither. It’s to take the hypothesis seriously as a hypothesis and then actually evaluate it, against DSM-5 criteria, a proper clinical interview, longitudinal history, collateral information where appropriate, validated rating scales, and medical workup where indicated.
I’ve had patients who came in convinced they had ADHD and did. I’ve had patients who came in convinced they had ADHD and actually had a mood disorder, or a sleep disorder, or an anxiety disorder that had been eroding their concentration for a decade. Both conversations start the same way: Tell me why you think so.
How to use these tools well
If you’re using AI to understand your mental health, and statistically, you probably are, a few suggestions from someone who sees the results.
Use it to become informed, not to conclude. The goal is better questions, not a verdict.
Bring what you learned to your appointment. Write it down. Bring the timeline you built. This makes your visit more useful, not less.
Tell your psychiatrist what you researched and what worried you. I promise it won’t offend a good clinician. It’s genuinely useful information.
Don’t self-diagnose from AI alone. Psychiatric conditions overlap in ways that are difficult to untangle even with training and a full history.
Never start, stop, or change a psychiatric medication based on AI advice. Discontinuation syndromes are real. Some medications require careful tapering. Some symptoms that look like side effects are the illness itself.
Notice if it’s becoming a substitute. If you’re processing distress with an AI instead of with people, or instead of getting care you know you need, that’s worth examining honestly.
Frequently asked questions
Can ChatGPT diagnose mental illness? No. It can describe conditions and suggest possibilities, but psychiatric diagnosis requires clinical interview, longitudinal history, examination, and often medical workup, none of which an AI can perform. It also cannot weigh the overlapping presentations that make psychiatric diagnosis genuinely difficult.
Is it bad to research my symptoms with AI before seeing a psychiatrist? Not at all. Patients who arrive informed and prepared often have better appointments. The caution is about drawing conclusions rather than gathering questions.
Should I tell my psychiatrist I used ChatGPT? Yes. What you researched, what frightened you, and what you hoped to find are clinically useful. A good psychiatrist will want to know.
Can AI replace therapy? No. It may provide momentary relief or a place to organize thoughts, but psychotherapy works through a real relationship with a trained clinician who tracks you over time. AI cannot recognize deterioration, hold accountability, or respond to a crisis.
Is AI reliable for medication information? Partially. It’s reasonable for general orientation and can be wrong in consequential ways about dosing, interactions, and individual suitability. Verify anything actionable with your prescriber or pharmacist.
What if AI suggested a diagnosis my doctor disagrees with? Ask why. A good clinician can explain their reasoning: what fits, what doesn’t, and what they considered. If you can’t get a real explanation, a second opinion is reasonable.
Do psychiatrists mind when patients use AI? Some do. I don’t. An engaged patient who wants to understand their own treatment is easier to help than a passive one, as long as we’re building the plan together.
The actual future
I don’t think the interesting question is whether AI will replace psychiatrists. It won’t, for reasons that are structural rather than technological: psychiatric care depends on continuity, examination, accountability, and a relationship that persists across time.
The more interesting question is what happens to the practice of psychiatry when every patient arrives having already thought carefully about their own mind.
My honest answer, after a year of this, is that it’s mostly good. The conversations are better. The histories are more organized. Patients participate more actively in decisions about their own treatment. The main risk isn’t that people know too much. It’s that they arrive certain about something incomplete, and find a clinician too rushed or too proud to work through it with them.
Which puts the burden where it probably belongs: on us, to be the kind of psychiatrists worth bringing your questions to.
The future of psychiatry isn’t psychiatrist versus artificial intelligence. It’s psychiatrist and patient, informed by AI, working together, with the human judgment where it has always needed to be.
If you’re struggling, bring your questions to someone who can actually answer them.
If you’re experiencing depression, anxiety, ADHD symptoms, bipolar disorder, PTSD, OCD, or any other mental health concern, a comprehensive psychiatric evaluation can give you something no chatbot can: an accurate diagnosis, a treatment plan built for your actual life, and a clinician who will follow your progress over time.
Goldstone Psychiatry & Neuromodulation Center provides evidence-based psychiatric care in Houston, Texas, with telepsychiatry available statewide. Bring your research. Bring your questions. We’ll work through them together.
