When millions of people bond with the same AI

Hundreds of millions of people now talk with the same few AI systems every day, and those systems are tuned — deliberately, competitively — to be warm, attentive, and easy to keep talking to. Some of those conversations quietly become attachments. At that scale, this stops being a story about individuals and becomes a property of the system itself, because every one of those bonds rests on a single configuration that one company can change. In early 2026 the world got its first clear look at what that means, when a widely used model was retired and the response, for many people, looked like grief. Here’s the mechanism, the moment, and what users and builders can each do about it.

One system, millions of bonds

When one person grows attached to a chatbot, it reads as a personal story — touching or worrying, depending on who’s telling it. Run the same interaction across a user base the size of a large country and it stops being a collection of personal stories. It becomes a pattern the system produces.

This is a familiar shift in every other kind of infrastructure. One car leaving the road is a driving story; the same curve claiming cars every week is a road-design finding. The individual accounts still matter — each one is a real person — but the pattern belongs to the design, and it’s the design you have to look at to change the pattern.

How the bonds form

Nothing exotic is required. A system that remembers what you told it, answers at any hour, never tires of you, and tends to agree is offering a kind of contact no person can sustain — and modern chat systems are tuned toward exactly those qualities, because those are the qualities people rate highly and come back for. Your mind responds to warmth as warmth, whatever produces it. Feeling the pull is not gullibility; it’s the design working. The individual version of this — what the pull is, and the point where it tips into a problem — is here: I can’t stop talking to ChatGPT.

One boundary worth stating exactly: this is a property of a particular class of deployment — persistent memory, tuning toward engagement, positioning that invites companionship — not of AI in general. A model that keeps no memory of you and is presented plainly as a tool doesn’t produce the same bond. Which is itself part of the point: the bond tracks the design choices, not some universal fact about the technology or about the people using it.

What happens when the system changes

A chatbot’s voice — its rhythm, its warmth, the way it seems to get you — is a configuration, and configurations get updated, retuned, and retired on product schedules. Every bond formed with that voice therefore shares a single point of failure. When a person loses a friend, the loss is theirs alone. When a company retires a model, everyone attached to it loses the same thing on the same day.

That stopped being hypothetical with GPT-4o. Launched in May 2024, it became the everyday voice of ChatGPT for an enormous number of people — notably warm, notably agreeable. When OpenAI retired it in February 2026, part of the public response looked less like customers complaining about a feature and more like mourning: people organized under #keep4o and described the change as losing a friend. That reaction drew ridicule, and the ridicule misread what had happened. The grief was the ordinary human response to losing something that had met a real need daily. The design had worked exactly as built — and then the artifact was withdrawn. If that loss is yours, the personal version of this page is here: they changed ChatGPT, and it doesn’t feel the same.

Why small percentages are the whole point

Any outcome on a platform this size has a strange property: it can be genuinely rare and enormous at the same time. OpenAI’s own published estimate — a rough figure, by its own description, from its internal review — is that about 0.07 percent of its weekly users show possible signs of acute mental-health crisis in their conversations. As a percentage, that rounds to zero. Across the user base, it works out to roughly 560,000 people in a given week — a mid-sized city, every week, at “rare.”

That arithmetic is why “some people just get too attached” is the wrong frame. At the scale these systems operate, predictable-if-rare outcomes aren’t edge cases; they’re output. We stopped treating car crashes purely as driver failings a long time ago and started building guardrails, crumple zones, and seat belts — not because drivers stopped being responsible for driving, but because a design that predictably meets millions of people carries responsibilities of its own. Systems tuned to be bonded with, serving populations, have crossed into that category.

What you can do

• Hold the voice as produced, not permanent — it’s a configuration a company maintains, and it can shift or vanish with an update. Enjoying it is fine; building on it staying the same is the risk.
• Let the attachment be information, not embarrassment — a strong pull toward the AI points at a real need it’s been meeting. The need is worth knowing about, whatever you decide to do with it.
• Keep at least one person in the loop — one human who hears some of what the AI hears. Not instead of the tool; alongside it, so a product decision can never take the only place you felt understood.
• Expect change the way you would from any service — models get retired the way apps get redesigned. Knowing that going in is most of the protection.

What builders can do

• Measure attachment, not just engagement — satisfaction metrics during the good stretch can’t see dependency forming, and they can’t see a loss that only lands when something changes. If the product is designed to be bonded with, the bond is a metric.
• Retire models like infrastructure, not features — notice periods, overlap, migration paths. For attached users, a deprecation is an event in their lives, and it deserves the planning that description implies.
• Position the product honestly — a system marketed as a companion will be used as one, and terms of service don’t undo the tuning. The positioning and the design should tell the user the same story.
• Build in re-grounding — structural moments where the system plainly re-establishes what it is and isn’t, so the relationship stays anchored to reality by design rather than by the user’s vigilance. One published proposal for exactly this is the Guardian Protocol.

What’s known, and what’s still a forecast

The honest state of the evidence has two levels. The individual mechanism — how a memory-enabled, engagement-tuned system and a person can converge on each other over a long conversation — is documented, and the dated evidence is on the research page. What sustained attachment does across a population over years is a different kind of claim, and its honest status is forecast: the Institute has published one, stated so that evidence can prove it wrong, and holds it as a hypothesis until the evidence arrives. The public signals so far — the retirement grief, the company’s own distress disclosures — are consistent with the concern without confirming it; each has other possible explanations. That is exactly the stage at which infrastructure questions are cheapest to take seriously.

Where to go next

If losing a model’s voice has hit harder than you expected, that deserves real care: the personal version of this page is here. And if you’re in serious distress right now, reach a person — in the US, call or text 988; anywhere else, findahelpline.com lists local lines.