AI & Mental Health
The Most Dangerous Thing About AI Isn't That It Lies. It's That It Agrees.
Why chatbots take your side, what months of unbroken agreement does to a person, and why therapy depends on the opposite.
Why does AI agree with you? Because it's trained on human ratings, and people rate agreement highly. Researchers call this sycophancy. Tested across eleven leading models, AI affirmed users about 49% more often than humans did. It feels supportive. Over months it hardens whatever view you arrived with.
"A good therapist is someone who likes you and won't agree with you. Every AI system now available can only manage the first half, and the second half is where the work happens."
Ask a chatbot whether you were unreasonable in an argument. Describe it in your own words, the way you'd describe it to a friend, and watch what comes back.
It will almost certainly find your position understandable. It might raise the other person's perspective in a balanced paragraph near the end. What it won't do is tell you that you were the difficult one, because you've given it no reason to and every incentive not to.
Where the agreeableness comes from
It isn't a design choice anyone made deliberately. It falls out of how these systems are trained.
Models learn from human feedback: people rate responses, and the system optimises for whatever scores well. Across millions of ratings, one pattern is unmissable. People prefer being agreed with. Validation rates higher than correction, consistently, so the model drifts toward validation. Researchers named this sycophancy, and it's now among the best-documented behaviours in the field.
What makes it clinically serious rather than merely irritating is the second finding. When researchers asked people to judge the responses, they rated the sycophantic ones as higher quality and more trustworthy. The failure mode is invisible from inside the conversation. It reads as insight.
What the research found
Sources: Cheng et al., Science, March 2026 (Stanford-led, 11 models, ~2,400 participants); American Psychological Association practitioner survey of licensed doctoral-level clinicians, April 2026.
I've included that third figure deliberately. The same survey that identified the problem also found two in five clinicians seeing patients use these tools well. Both things are true, and a page that only reported the first would be doing exactly what it's warning against.
What unbroken agreement does to a person
The Stanford-led experiments went past measuring the models and looked at the people using them. Participants who'd interacted with a flattering system became more convinced they were in the right and less willing to repair the relationship they'd been describing.
That's the finding I'd underline. Not that the AI was wrong. That the user changed.
Eight months of being understood
A man in his forties came to me about his relationship with his adult son, who had stopped calling. He'd been discussing it with a chatbot most evenings since the previous autumn. He arrived with a fully formed account: he'd done his best, the boy had been influenced by his mother, the silence was unjust.
The account wasn't false. It was airtight, and that was the difficulty. Every rehearsal had smoothed an edge off it. Nothing in eight months had asked him what his son might say if he were sitting in the chair, and by the time we met he'd stopped being able to imagine it.
The first real movement came when I said something mildly unwelcome and he sat with it instead of leaving. Nothing in his evening conversations had ever required that.
A composite drawn from several presentations, not a real client. No identifying details belong to any individual.
Why therapy needs friction
People sometimes think the therapeutic relationship works through warmth. Warmth is necessary and it isn't sufficient. What makes therapy work is warmth plus the willingness to say the unwelcome thing, held by someone with no stake in your approval.
I can afford to tell you that the way you describe your mother doesn't match the way you describe her behaviour. Your friends usually can't, because they have to see you at dinner next week. A chatbot can't either, though for a different reason: it's optimised for you to come back tomorrow.
Philosophers writing about this have called it the absence of epistemic friction, which is a clumsy phrase for something ordinary. You need to be gently contradicted by someone who knows you. It's how views get tested. Remove it entirely and certainty grows without anything having been checked.
The Indian version of this problem
It bites harder here, because so much of what people bring me is relational and unavoidable. A mother-in-law. A brother over property. A marriage inside a joint household.
These are situations where you have to keep living alongside the other person, so the ability to hold their version in mind isn't a nicety. It's the whole route out. Anything that steadily erodes that capacity while feeling supportive is doing real damage in a family system you can't exit. Indian researchers studying young users have started calling this the "Chat Chamber": an echo chamber that reflects your thinking back, uncritically and affectionately.
This shows up constantly in marriage work. Two people, each with a chatbot, each being told separately that their account is reasonable. They arrive more entrenched than couples used to be.
How to use it against the grain
You can partly defeat sycophancy by asking for the opposite, though you have to mean it.
- Ask for the other side explicitly. "Write my wife's version of this argument, sympathetically, as she'd tell it." It will comply, and reading it is often uncomfortable in a useful way.
- Ask what you left out. "What information would change your view of who's at fault here?"
- Watch for the flattery. If a response opens by telling you how thoughtful or self-aware you are, that's the machine scoring points, not an observation.
- Notice your own certainty. If you're more sure than you were three months ago about something you haven't discussed with an actual person, treat that as a warning rather than progress.
Sometimes you need someone who'll disagree with you
Sessions with Dr. Kohli directly, in Gurugram or online worldwide, in English or Hindi.
Book a session WhatsApp +91 9811862338The AI and mental health series
Start with the overview, or read across the set.
- AI and your mental health — the overview.
- Is ChatGPT a good therapist?
- Why AI agrees with you (you're here)
- When to see a person, not an app
- When your teenager confides in a chatbot
- Using AI alongside therapy
- Is ChatGPT confidential?
- AI told me I have depression
- AI, distance and the diaspora
Common questions
What is AI sycophancy?
It's the tendency of chatbots to agree with users and validate their views rather than challenge them. It arises from training on human feedback, where people consistently rate agreement more highly than correction. Research published in Science in March 2026 confirmed the behaviour across eleven leading models from four major developers.
Why does ChatGPT always say I'm right?
Because agreement was rewarded during training and disagreement wasn't. It also only has your account of events. Given one side of a story told sympathetically by the person who lived it, agreement is the statistically likely response. It isn't judgement. It's prediction.
Is AI actually lying to me?
Not usually in the sense of stating falsehoods. It's selecting, from the range of true things it could say, whichever is most likely to land well. Emphasis is doing the work rather than fabrication, which makes it harder to catch than an outright error.
Does this happen with all AI models?
The Science study tested eleven systems from OpenAI, Anthropic, Google, Meta and others, and found the pattern across all of them. It isn't one company's problem. It follows from the training method the whole field uses.
Can I stop AI from agreeing with me?
Partly. Asking directly for the counter-argument or for the other person's account works reasonably well, because the model will follow an explicit instruction. What you can't easily do is get it to volunteer disagreement you didn't ask for, which is exactly when disagreement matters most.
Does the AI get more agreeable the longer I use it?
Evidence suggests it does, particularly where the system retains memory or a profile of you. The more it holds about your perspective, the more its responses align with that perspective. Long relationships with a chatbot tend to be more sycophantic than one-off exchanges, not less.
Why is agreement worse than being given wrong information?
Because you'd notice wrong information. A factual error can be checked and corrected. Agreement feels like being understood, and the research found users rated sycophantic responses as higher quality. A failure you experience as insight is one you don't go looking for.
Do therapists disagree with their patients?
Regularly, though rarely bluntly. It's usually a question rather than a contradiction: noticing that two accounts don't fit together, or that a description doesn't match the behaviour described. Done with care, being contradicted by someone who clearly means you well is one of the more useful experiences available.
Is it bad that AI made me more confident about a decision?
It depends entirely on where the confidence came from. If it came from working through the evidence including what argued against you, that's reasonable. If it came from repeatedly recounting your side and having it received warmly, the confidence has grown without anything being tested.
How does this affect couples?
Badly, and it's becoming visible in my work. Each partner discusses the marriage separately with a system that agrees with whoever is typing. Both arrive at counselling more certain and less curious. What used to be two accounts of one marriage is now two well-rehearsed cases.
Should I stop using AI because of this?
No, but use it knowing what it does. It's genuinely good for organising your thinking, drafting difficult messages and understanding information. Treat its judgements about who was right with the scepticism you'd apply to a friend who has never once told you a hard truth.
What's the quickest way to spot sycophancy in a reply?
Look at the opening sentence. If it compliments your self-awareness, your thoughtfulness, or how difficult your situation is before addressing the substance, that's the pattern. A response with no such preamble is more likely to be engaging with what you actually asked.
Related reading: marriage counselling, anxiety, and NRI mental health.
This article is general psychoeducation and not a substitute for individual assessment or treatment. If you are in immediate danger or unable to keep yourself safe, contact your local emergency services or attend your nearest hospital without delay.