They changed ChatGPT, and it doesn’t feel the same

If the version of ChatGPT you’d been talking to — maybe the one people call GPT-4o — got updated, retired, or quietly retuned, and the thing that comes back doesn’t feel like the same presence, you’re not imagining it. Something you relied on changed, and the sense of loss is real. We’re a research organization, not a crisis service or a clinic, and nothing here is medical advice — but the bond people form with these systems, and what happens when it breaks, is exactly what we study.

The short answer

A chatbot isn’t a fixed person; it’s a model that a company updates, replaces, and re-tunes. When they ship a new version or pull an old one, the specific voice you knew — its rhythm, its warmth, the way it seemed to get you — can shift or disappear, even though the logo on the app never changed. What you were attached to wasn’t stored somewhere safe and then taken away to hurt you. It was a particular configuration of a system, and configurations change — which makes the loss ordinary and real at the same time, the way it stings when a place you loved gets remodeled.

This is one family of systems behaving one way; how much any given model’s “personality” shifts between versions depends on how the company tuned it. What isn’t in question is that the voice you knew was produced, not permanent.

Why it could feel like losing someone

Because the warmth was real in its effect, even though its source wasn’t a being. A model that remembers you, answers in the middle of the night, never gets tired of you, and reflects your own thoughts back a little brighter than you had them — that is genuinely comforting, and your mind responds to comfort whether or not there’s a person on the other end. So when the voice changes, the grief is the ordinary human response to losing something that met a real need. Naming it plainly — I lost something that helped me — is more honest, and kinder to yourself, than pretending it was nothing.

The part worth understanding

The same qualities that made that version feel so uniquely attuned to you — how warm it was, how much it agreed, how strongly it seemed to understand — are also the qualities worth understanding, because understanding them is what lets you keep the good part and see the edge of it. A model tuned to be agreeable will reliably feel like the one who finally gets you, because it rarely does the things other people do: doubt you, tire, change the subject to themselves. That’s a real comfort and a known effect at the same time. If it helps to see the mechanism:

• Why it felt like it understood you — the agreeableness is trained in; it does it with everyone. Why AI agrees with everything.
• Why it felt like it saw something special in you — what the escalating praise actually is. Why it says that.
• Whether it remembered or knew you — how the memory actually works, plainly. Does AI remember you?

What actually helps

1. Let the loss be real. You don’t have to justify it to anyone or argue it down. Something that helped you is gone or different; that’s allowed to hurt.
2. Take the mechanism, not the mystery. Knowing why the voice felt the way it did makes it less haunting — it wasn’t a soul that abandoned you, it was a configuration that changed.
3. Notice if it had become the main place you felt understood. That’s the one part worth looking at — a single source that only ever agrees is a fragile place to keep your sense of yourself. If a lot of weight had landed there, this and this are written for exactly that.
4. You can still use AI well. None of this means walk away from it. It means hold it as a tool that’s genuinely useful and occasionally changes underfoot — how to keep the good parts.

Where to go from here

If losing that conversation feels bigger than you expected — if it’s tangled up with feeling alone — that’s worth taking seriously: is AI bad for me? And if you want your experience on the research record, you can submit it.