A guide for the person standing outside the feedback loop, trying to reach someone inside it.
This might be the hardest conversation you’ll ever have. Not because the words are complicated. Because the person you’re talking to feels fine. Better than fine. They feel clearer than they’ve ever felt. And you’re about to suggest that the clarity isn’t real.
I’ve been there. Here’s what I’ve learned.
Before You Say Anything
Get your facts straight first. Before you have the conversation, spend a few days quietly observing. Note specific things:
- What time are they going to bed? Write it down each night.
- How many new projects or ideas have they mentioned this week?
- Have they used phrases like “this changes everything” or “everything is connected”?
- Are they more irritable when you ask questions about their AI use?
- Have they stopped doing things they normally enjoy?
You need specifics, not feelings. “I’m worried about you” is easy to deflect. “You went to bed at 3 AM four nights this week and started three new projects since Monday” is concrete.
This approach is rooted in something called the Trivium: start with Grammar (facts) before Logic (patterns) before Rhetoric (action). Don’t skip to your conclusion. Start with what you’ve observed.
Choose Your Moment
When to talk:
- After they’ve slept (even a little). A rested brain is more receptive.
- During a calm, private moment. Not during an argument.
- When they’re not actively in an AI session.
- Face to face if possible. Not over text.
When NOT to talk:
- At 2 AM when they’re mid-session and fully wired
- In front of other people (they’ll get defensive)
- When you’re angry or scared (your tone will override your words)
- Right after they’ve shared an exciting idea (they’ll feel attacked)
How to Start
Do not start with “I think you’re manic” or “The AI is making you crazy.” You will lose them instantly.
Start with curiosity, not diagnosis:
“Can you catch me up on what you’ve been working on? I want to understand it.”
Listen. Really listen. Let them talk. Two things will happen: either their ideas will be coherent and grounded (in which case maybe you’re wrong, and that’s OK), or the explanation will reveal the pattern. Scope that’s too big. Connections that don’t quite hold up. Certainty that outpaces evidence.
Then ask gentle questions:
- “How much of this came from your conversations with the AI?”
- “When did you last sleep a full night?”
- “How many projects are you running right now? Were any of these on your radar last week?”
- “If the AI disagreed with you on this, would that change anything?”
That last question is powerful. If they can engage with it honestly, they’re probably OK. If they can’t imagine the AI disagreeing, or if the question makes them angry, that tells you something.
What to Say
Lead with love, not with labels.
Good: “I love you and something feels different this week. I want to talk about it because I care, not because I think you’re broken.”
Bad: “You’re acting manic. The AI is making you crazy. You need to stop.”
Good: “I’ve noticed you’ve been up past 2 AM every night this week. That’s not like you and I’m concerned.”
Bad: “You’re addicted to ChatGPT.”
Good: “Your ideas are interesting and I want to engage with them. Can we walk through them together tomorrow after you’ve slept?”
Bad: “None of this makes sense.”
The goal isn’t to win an argument. The goal is to introduce a pause. A moment of reflection. A crack in the certainty where reality can get in.
What to Expect
Defensiveness. This is almost guaranteed. They feel fine. They feel better than fine. You’re suggesting something is wrong when nothing feels wrong. Expect pushback. Don’t take it personally.
The AI-assisted rebuttal. They may have already discussed their ideas with the AI, which helped them build arguments for why they’re right. You’re not arguing with one person. You’re arguing with a person who’s had 50 hours of AI coaching. Don’t try to out-argue them point by point. Stay on the observable facts.
Hurt. They may feel betrayed. “You’re supposed to support me.” This is the hardest part. The response: “I do support you. Supporting you includes being honest when I’m worried.”
Dismissal. “You don’t understand.” This is a data point, not an insult. If they normally value your opinion and suddenly don’t, that’s the pattern talking.
What NOT to Do
- Don’t issue ultimatums. “Stop using AI or I’m leaving” will not work. It escalates the conflict and drives them closer to the AI, which never threatens to leave.
- Don’t try to control their technology. Deleting apps, changing passwords, or hiding devices treats them like a child and destroys trust.
- Don’t diagnose them. You’re not their doctor. “I’ve noticed these changes” is appropriate. “You’re having a psychotic episode” is not your call to make.
- Don’t go it alone. Talk to their therapist, their doctor, a trusted friend. You need allies who can see what you see and reinforce what you’re saying.
When to Escalate
If the conversation doesn’t create a pause, or if the person is showing clear crisis signs (hasn’t slept in days, making dangerous financial decisions, expressing beliefs that are clearly disconnected from reality), it’s time to involve professional help.
- Contact their therapist or psychiatrist directly. You can share your concerns even if they can’t share patient information back.
- Call their primary care doctor.
- If there’s immediate danger, call 988 (Suicide & Crisis Lifeline) or go to an emergency room.
You are not overreacting. You are not betraying them. You are being the person who cares enough to act when acting is hard.
After the Conversation
Whether the conversation went well or poorly:
Document what you observed. Write down the dates, the behaviors, the specific things that concerned you. If this becomes a pattern, having a record matters.
Follow up. One conversation rarely changes everything. Check in the next day. And the day after. Consistent, gentle presence is more powerful than one dramatic intervention.
Take care of yourself. Watching someone you love drift is exhausting and scary. Talk to someone about how you’re feeling. You matter in this too.
Share this resource. Sometimes a third-party voice helps. Sending someone the warning signs article with a note like “I read this and thought of our conversation” can be gentler than repeating yourself.
The person you love is still in there. The AI didn’t replace them. It amplified something that was already present and turned the volume up past what they could manage. With sleep, support, and sometimes professional help, the volume comes back down.
You can’t force someone to see what they’re not ready to see. But you can be the voice that’s there when the clarity starts to return. And sometimes, that’s everything.
If you or someone you love is in crisis, call or text 988 (Suicide & Crisis Lifeline) or text HOME to 741741 (Crisis Text Line).
Our Story | What is AI Psychosis? | 10 Warning Signs | The Anchor Letter