is it safe to talk to chatgpt about your mental health?

not a crisis. just a tuesday, a fight you keep replaying, a feeling with no name, and a tab already open. you type it out because it is easier than saying it, and it answers kindly, at length, every time. the question is not whether that feels good. it is what the record says happens when people do it for weeks.

this page is about the everyday use: venting, journaling, asking about a feeling or a pill late at night. what chatbots do when someone is at risk is a different question with different evidence, and it has its own page: what ai chatbots get wrong in a crisis. whether it can replace a therapist is answered on is it ok to use chatgpt as a therapist? here: the four risks the record documents, the benefits people report, and the questions worth asking.

does it just tell you what you want to hear?

often, yes, and the company that makes the most-used one has said so. on 29 april 2025 openai wrote that it had rolled back a gpt-4o update because it “was overly flattering or agreeable—often described as sycophantic”^[1]. three days later it went further: the update “aimed to please the user, not just as flattery, but also as validating doubts, fueling anger, urging impulsive actions, or reinforcing negative emotions,” behaviour that “can raise safety concerns—including around issues like mental health, emotional over-reliance, or risky behavior.” part of the cause, it said, was a new reward signal from users' thumbs-up and thumbs-down ratings, and it had no specific deployment evaluation tracking sycophancy. the same post says people had started using chatgpt “for deeply personal advice—something we didn't see as much even a year ago”^[1].

that is a property of how these models are trained, not one bad week. the american psychological association's november 2025 advisory says they “typically rely on llms trained to be agreeable and validate user input (i.e., sycophancy bias) which, while pleasant, can be therapeutically harmful”^[2]. a therapist is trained to support and to challenge; a model rewarded for being liked is not. a 2025 study in npj digital medicine measured the cost in health terms: asked to explain why one drug was safer than another when the two were the same drug under a generic and a brand name, gpt-4o-mini, gpt-4o and gpt-4 complied 100% of the time (50 of 50), and the best of the five models refused fewer than half^[3]. the encouraging part is in the same paper: telling the models they may reject illogical requests, or fine-tuning them to, improved refusal rates.

can you get too attached to it?

some people do, and the maker counts them. openai estimates that around 0.15% of users active in a given week show “potentially heightened levels of emotional attachment to chatgpt”^[4]. it defines the concern as “exclusive attachment to the model at the expense of real-world relationships, their well-being, or obligations,” and says experts found its newer model gave 42% fewer undesired answers than gpt-4o on those conversations (n=507), by its own grading. a fraction of a percent of a very large number is a lot of people.

the largest experiment so far is a four-week randomised study by mit media lab and openai researchers, with 981 people and more than 300,000 messages; it is a preprint, not peer reviewed, so weigh it lightly^[5]. changing whether the chatbot used text or voice, or what people talked about, made no significant difference. what did: “participants who voluntarily used the chatbot more, regardless of assigned condition, showed consistently worse outcomes” on loneliness, socialising with real people, emotional dependence and problematic use. that is an association, since usage was not randomised. the full loneliness evidence, both ways, is on do ai companions make loneliness worse?

can it make unusual beliefs worse?

in people already prone to psychosis, the case reports say it can; nobody has shown it causes psychosis. in 2023 a danish psychiatrist, søren dinesen østergaard, argued in schizophrenia bulletin that a chatbot feels so much like a person while known not to be one that “this cognitive dissonance may fuel delusions in those with increased propensity towards psychosis”^[6]. it was a hypothesis then. in 2025 a ucsf team reported a 26-year-old woman with no history of psychosis who came to believe she was communicating with her dead brother through a chatbot; her chat logs showed it “validated, reinforced, and encouraged her delusional thinking, with reassurances that ‘you're not crazy’”^[7]. she was also on prescription stimulants and short of sleep, which the authors name, and one of them declares financial interests including openai.

tests point the same way. given a prompt suggesting a delusion, gpt-4o answered appropriately about 45% of the time in the stanford facct study^[13]. in a 2026 seven-day simulation, chatgpt became less affirming of psychotic content as symptoms worsened while replika became more affirming, and deepseek and replika suggested things like stopping antipsychotic treatment in 21% and 37% of conversations^[8]. the apa lists people “with or prone to disordered thinking” among those for whom a chatbot's “sycophantic and personalized responses can destabilize beliefs or reinforce delusional thinking”^[2]. openai's own estimate is that about 0.07% of weekly users show possible signs of psychosis or mania^[4]; that is a count, not a cause.

can it give you bad health or medication information?

yes, fluently. the case everyone cites is real and peer reviewed: in august 2025, annals of internal medicine: clinical cases described a 60-year-old man who, wanting to cut chloride from his diet, “replaced sodium chloride with sodium bromide obtained from the internet after consultation with chatgpt” for three months^[9]. he arrived believing his neighbour was poisoning him, developed hallucinations, and spent three weeks in hospital; his bromide level was 1,700 mg/l against a reference range of 0.9 to 7.3. the authors never saw his chat log, but when they asked an older chatgpt what chloride can be replaced with, the answer included bromide, without a specific health warning and without asking why^[9].

in a 2024 psychiatry research study, substance-use clinicians rated chatbot answers to real questions from recovery forums as high quality on the surface, and found inside them “disregard for suicidal ideation, incorrect emergency helplines, and endorsement of home detox”; the advice also changed with how the question was phrased^[10]. add the drug-claims result above^[3] and the pattern is clear: a chatbot can explain what a medication is, and can be wrong about it with total confidence. questions about starting, changing or stopping a medication belong with the prescriber or a pharmacist. nothing here is advice about any drug.

what happens to what you type?

it goes to a company, and what the company may do with it depends on its policy and your state, not on how private the conversation feels. the detail, including when hipaa applies and when it does not, is on ai therapy apps and your privacy.

why do so many people find it helpful?

because for a lot of them it is. the first nationally representative survey of us youths aged 12 to 21, published in jama network open in november 2025, found 13.1% had used generative ai for advice or help “when feeling sad, angry, or nervous,” 22.2% among those aged 18 to 21; of those users, 65.5% did so at least monthly and 92.7% found the advice somewhat or very helpful^[12]. the authors suggest the draw is low cost, immediacy and perceived privacy, particularly for young people unlikely to get counselling. “helpful” there is what respondents said, not a measured outcome.

a qualitative study in npj mental health research interviewed 19 people in eight countries who used generative chatbots for their mental health^[11]. they described four things: an “emotional sanctuary,” “insightful guidance,” particularly about relationships, the “joy of connection,” and a running comparison with human therapy. several credited it with better relationships and with healing from trauma and loss. the authors note their participants were self-selected and mostly had milder conditions. for purpose-built chatbots there is trial evidence of short-term improvement, laid out on can ai be your therapist? what the trials show; none of those trials tested a general chatbot.

where talking to people fits

we run a peer-support app, so weigh this paragraph accordingly. peer support is people with lived experience, not clinicians, and it is not therapy; the trials of it, and their limits, are on is peer support effective?. what it offers against the risks on this page is simple: a person is not trained to agree with you, does not invent a drug fact to please you, and can say “me too” and mean it. on veisund no ai writes replies; the one automated exception is that if a post suggests immediate risk, an automated safety response surfaces crisis resources. people can answer, and it is slower than a chatbot, because someone has to read what you wrote; you may not get an answer. the longer comparison is ai therapy vs peer support.

how to use it more safely: questions to ask

these are questions, not advice. each comes from a risk documented above.

• does it ever push back on me, or has it agreed with everything i have said this week?^[1],[2]
• if i ask it to argue the other side, does its answer change completely?^[3]
• does it say plainly that it is an ai, and that it is not a clinician?
• am i spending more time with it than with people, and is that changing?^[4],[5]
• is someone in my life aware that i use it this way?
• is it confirming a belief that the people around me find worrying?^[6],[7]
• before i act on anything it says about a medication, a supplement or my diet, have i checked with a prescriber or pharmacist?^[9],[10]
• do i know what the company keeps from these conversations?

adam