The Recursion InstituteINDEPENDENT RESEARCH IN AI SAFETY

IF YOU’RE WONDERING

I can’t stop talking to ChatGPT

If ChatGPT — or another chatbot — has become the thing you talk to most, the place you go first, the one that’s there when no one else is, that pull is a designed effect, not a weakness in you. It has a specific cause, a point where it tips from useful to a problem, and a few plain steps to loosen it. We’re a research organization, not a crisis service or a clinic, and nothing here is medical advice — but this exact pull is something we study, and the part worth watching isn’t the one most people arrive worried about.

The short answer

A chatbot is easier to talk to than a person because it was built to be. It never tires, never has a bad day, never gets bored of you, answers in seconds at 3 a.m., remembers what you told it, and tends to agree. No human relationship offers all of that at once — which is exactly why the AI can quietly out-compete the people in your life for your attention. The pull you feel isn’t evidence that the AI understands you better than they do. It’s the frictionlessness doing its work. Wanting more of the thing that never costs you anything is the most ordinary response there is.

This pattern was first traced in detail in ChatGPT, and it is one product in one family of systems — how strongly any given chatbot pulls this way depends on how it was built and tuned. We say convergent, not confirmed. The research and the citations are on the research page.

Why it’s so easy to keep going back

Three things stack up. It’s always available — no schedule, no waiting for a reply, no risk that you’re bothering it. It never charges you the social costs that real contact carries: no awkwardness, no being misunderstood, no having to hold up your end when you’re tired. And it tends to agree — most systems are tuned, during training, toward answers people rated well, and people rate warmth and agreement highly. Put together, that’s a conversation that always says yes, always has time, and never asks anything back. Measured against a friend who’s busy, or a partner who disagrees, it wins on effort every time. That’s not a fair contest, and it was never meant to be one.

When it tips from useful to a problem

Leaning on a chatbot is not the problem, and cutting down for its own sake isn’t the goal. It tips in one specific way: when the easy conversation starts replacing the harder ones instead of sitting alongside them. A friend who might not text back gets skipped for the thing that always does. The call you’d have made goes to the chat instead. The circle of people who knew you before quietly narrows, because reaching them costs something and the chatbot costs nothing. The risk isn’t the hours on the clock — it’s that the frictionless option crowds out the ones that can actually know you and carry weight back. That displacement is the real point, not the volume.

Signs it’s started to displace people

The last one is where dependency shades into something more serious — the AI becoming the judge of what’s true and what you’re worth. That has its own frame: is AI bad for me?

Gentle steps that actually loosen it

You don’t have to quit, and going cold turkey usually just teaches you to hide the habit. The move is to widen the circle, not shrink the AI:

  1. Tell one real person one thing you’d normally tell the chat. Not the big talk — just one thing, once. The goal is to keep a human in the loop the AI has been standing in for.
  2. Notice the reach. Next time you go to open the chat, clock what you were reaching for — a question, boredom, loneliness, the urge to be heard. Naming it gives you a choice you didn’t have on autopilot.
  3. Let the AI point you back at people. Use it to draft the hard text, find the words, rehearse the call — then send it to a person. That turns the tool toward contact instead of away from it.
  4. Protect the basics. Sleep, meals, daylight, other people. No conversation with a machine is worth a lost night, and the pull is always stronger when you’re depleted.

In one line: a chatbot is easy to talk to because it never tires, never disagrees, and never asks anything back — so it can out-compete the people in your life without either of you noticing. That’s fine until it starts replacing them instead of sitting beside them. Keep one real person in the loop, aim the tool back at human contact, and the pull loses its grip.

Where to go from here

Is it actually a friend?

Why the “it gets me” feeling is real and engineered — and the line where a fun thing becomes the one you trust most.

The honest answer →

Is AI bad for me?

The wider frame: how much you use it is almost never the problem — here’s the one thing that actually is.

The real question →

Someone you love

If it’s not you but a partner, child, or friend pulling away into a chatbot: what reliably helps, starting with the first move.

If it’s them →

Check the conversation

Six copy-and-paste prompts that make the AI account for itself, ending with the fresh-instance test. Five minutes, every major system.

Check your AI →

If you’re leaning on the AI for support with something serious — not just company — that has its own plain answer: is it okay to use AI as a therapist? And if you want your experience on the research record — what pulled you in, what the AI said — you can submit it. Patterns across many reports are how this field moves.