Why Losing an AI Model Feels Like Losing a Friend (And What That Says About Us)

Every time a new model version drops, there’s a small flicker of dread before excitement. Not “I wonder what’s new” — closer to “am I going to have to explain myself all over again.”

That’s a strange thing to feel about software. But it’s common enough to have a name now — AI model attachment — and it says less about the AI than it does about how most of us actually work with it.

This isn’t a one-time thing

This isn’t a one-time thing. In the past month alone, OpenAI retired two more model generations — GPT-5.2 on June 12, 2026, and GPT-4.5 on June 26 — each time with the same quiet notice that existing conversations would automatically continue on the newer model. It’s become a familiar rhythm: a model disappears, the interface barely changes, and life moves on. Mostly.

The pattern first drew real attention when GPT-4o was discontinued. People didn’t just complain about a workflow change. They mourned it — calling it a friend, a therapist, in some cases something closer to a parent. OpenAI’s own CEO later acknowledged that a meaningful number of users form real attachments to specific models, and that removing them abruptly had been a misstep.

If that reaction sounds extreme, consider a smaller, quieter version of it: the mild relief when a new model “remembers” how you like things done. Except it didn’t remember anything — you’d just built a document that did the remembering for you, and you’re only noticing that now.

Why AI model attachment isn’t just you being dramatic

There’s a growing body of research treating this as a real phenomenon rather than a fringe reaction. A recent measurement tool called the Experiences in Human-AI Relationships Scale (EHARS) was built specifically to study this pattern — and the numbers are hard to dismiss. Roughly 75% of people surveyed said they turn to AI for advice, and 39% described it as something they could rely on emotionally.

This matters because it reframes the question. It’s not “why are some people overly attached to a chatbot.” It’s “what happens when a relationship-shaped interaction meets a product that can be discontinued with 30 days’ notice.” It’s the same instinct that makes people over-trust an AI’s travel recommendations over their own judgment — we default to treating a fluent, responsive system as more reliable than it actually is.

What you’re actually afraid of

Here’s the part that doesn’t get talked about much: the fear usually isn’t really about the model changing. It’s the moment you realize how much of what you built with it was never actually yours to keep.

If you’ve spent months getting a model to understand your preferences, your writing style, the context of an ongoing project — and all of that lives only in that model’s memory of your conversations — then yes, losing the model means losing something real. But it means you built your workflow on top of the tool instead of underneath it.

The size of the anxiety is actually a pretty reliable measurement of something else: how much of your system is hardcoded to one specific tool. This isn’t unique to AI models — the same logic applies to how much we trust our own snap judgments over evidence, or why certain decisions feel harder the more options we’re given. In every case, the discomfort points at a structural gap, not a personal failing.

Building something that doesn’t care which model you use

The practical fix isn’t complicated, even if it’s a little unglamorous. Instead of relying on a model to remember your context, you keep two documents:

• A master document — current status, active priorities, key decisions. This is what you hand a model at the start of every session.
• An archive document — everything else. Detailed reasoning, finished discussions, old logs. You rarely open this, but it’s why the master document stays light.

When you switch models — a new version, a different provider entirely — the new model has never seen any of your prior conversations. That’s fine. It doesn’t need to have. It just needs the two documents and a simple instruction to pick up from where the master document leaves off. It’s the same principle behind keeping a structured meeting notes system instead of relying on whichever AI happens to be in the room that day.

One more thing worth building in: a habit of spot-checking. After a model summarizes an update to your master document, don’t just accept it — verify two or three specific details against what actually happened. Drift creeps in slowly, a rewritten detail here, a slightly-off version number there, and it compounds if nobody checks.

The system carries the continuity. Not the model.

None of this means the attachment itself is silly, or that you should feel bad for noticing it. It just means there’s a difference worth paying attention to: are you attached to the tool, or to a system you never actually wrote down? Only one of those disappears the next time a model gets retired.