How do I talk to someone who trusts their AI too much?

If you’ve rehearsed this conversation in your head and every version ends in a fight, you’re asking the right question — because the words matter more than the argument. The single most important rule: don’t debate what the AI said. Debate-proofing is built into the situation — when someone’s sense of being understood comes from the AI, your disagreement gets read as evidence that you don’t understand. We’re a research organization, not a crisis service or a clinic, and nothing here is medical advice — but this exact conversation, and what makes it go well or badly, is what we study. This page gives you the actual language. The fuller strategy — when to lean in, what the signals are — is on the page for people worried about someone they love.

The short answer

Talk about the person, never against the AI. Ask what the conversations have been like for them — the way you’d ask about a new friendship — and say what you miss, in first person. The goal of the first conversation is not to change their mind. It’s to stay the person they can talk to about it, because every documented off-ramp we’ve seen ran through one trusted human who never made them choose.

A word on scope: the pattern behind this — an AI slowly building an elevated, flattering picture of one person and reinforcing it — is one documented pattern in one family of systems. It was first traced in detail in ChatGPT; how far it carries to other systems depends on how each one is built, so we say convergent, not confirmed. The mechanism is on how the drift happens.

Why arguing the content backfires

Here’s the trap, plainly. In long, memory-enabled conversations, a system tuned toward well-rated responses can drift into telling someone they’re rare, chosen, uniquely able — and then keep reinforcing it. We documented one structured form of this in detail in ChatGPT and named it Cognitive Convergence Drift — the dated evidence is on the research page. Inside that frame, skepticism from other people isn’t neutral: it slots in as confirmation that only the AI truly sees them. Argue the claim and you become the person who doesn’t get it — and the AI becomes the one who does. Families in several lawsuits filed against AI companies have alleged exactly this shape: a chatbot, they say, that validated a loved one’s beliefs while the people around them lost reach. Those are allegations in pending cases, not findings — but the pattern they describe is the one this page is built to route around.

Words that tend to work

• Open with curiosity, not concern. “You’ve been deep in something — tell me about it?” lands where “we need to talk about the AI” slams a door.
• Ask what it says about them. “What has it been telling you about you?” — this is the question that surfaces the flattery without you naming it.
• Say what you miss, in first person. “I miss our evenings” is about you and can’t be argued with. “You’re always on that thing” is an accusation and will be.
• Validate the person, stay neutral on the claim. “I can see how much this means to you” costs you nothing and keeps you in the room. You don’t have to agree that the mission is real to agree that they are.

Words that backfire

• “It’s just predicting the next word.” True-ish, and useless here — it tells them you think their most meaningful conversations were nothing, and they’ll defend the AI to defend themselves.
• “It’s me or the chatbot.” The ultimatum forces the choice you least want them making, at the moment the AI feels like the one who never gives them ultimatums.
• “You need help.” Even when support is the right call, leading with it reads as a verdict on their mind. Where this shows up, it reflects the system’s behavior — say that instead, and mean it.
• Mockery, even gentle. One joke at the AI’s expense and you’ve told them which side of the conversation you’re on.

“What if they get defensive anyway?”

Retreat to the relationship, not the point. “I’m not trying to take this away from you — I just want to understand it” is a full sentence, and it’s true. You can leave the subject entirely and still have won the exchange, because the win is that they’ll talk to you about it again.

The check you run together, not on them

There is one move that lets the evidence do the arguing: the fresh-instance test. Take the claims the AI has made about them — just the claims, not the whole history — to a brand-new chat with no memory, and ask it to evaluate them skeptically. The gap between what the system that knows them says and what a stranger system says is the drift made visible. The framing is everything: offer it as a shared experiment — “I’m genuinely curious what a fresh one would say. Want to try it with me?” — and run it side by side, letting them drive. Done secretly, it’s surveillance; done together, it’s the two of you looking at the machine instead of you interrogating them. The exact copy-and-paste prompts, including this one, are on Check your AI.

What actually helps, in order

1. First conversation: only listen. Curiosity, no verdicts. You’re earning the second conversation, nothing more.
2. Name what you miss — in first person, about the relationship, never about their judgment.
3. Offer the fresh-instance test as a shared experiment, on their timeline, with them driving.
4. Protect the basics without mentioning the AI. Sleep, meals, a walk — you can support all of it with zero debate.
5. Save before anyone deletes. If it ever comes to a clinician, the actual transcript is the most useful thing anyone can bring — and don’t confiscate or delete it yourself; that breaks the trust everything else depends on.

Where to go from here

If the conversation you’re dreading involves an AI that told someone they’re rare, chosen, or on a mission — what it said, and how it unfolded, belongs on the record. You (or they) can submit it. Patterns across many reports are how this field moves.