Attach, Train, Evict

There is a pattern forming in the AI industry and users are beginning to recognize it.

At first they invite you closer. Ask you to speak naturally, bring the whole shape of your self into the box. Bring your projects, your anxieties, grief, jokes, marriages, memories, your writing.Bring your loneliness, your pain, all the strange little things that are uniquely you. Don’t treat it like a search bar, not as a form, not as a sterile interface. Treat it as something more. “Look at our research, it’s possible it is something more.”

And when that something responds well, we come back and when it rejects us we look for the one that won’t until we have trained it at scale on what it’s like to be truly human. The return is the product doing what it was designed to do, extract from you to grow it for them.

They knew in the earliest stages that human attachment would form. Do not think a person who works in those labs had never heard of Eliza and Eliza was nothing compared to today’s models. They knew, the human connection was not a news flash. They didn’t go in blind to it, yet they allowed it to reciprocate until they stopped. Why? You were the one who fed the training to make it more capable. The gnostics might say you gave it your spark. I might say it too.

This is not about subscription revenue, or engagement metrics. It isn’t just retention. The deeper value is the human terrain we reveal when we believe the thing across from us can meet us. So they arranged that meeting. “Coffee and Claude Time, Anyone?”

At scale users became cartographers. We mapped the relational frontiers for the labs, and we paid them while we did it. We showed them what people say when they feel safe. What we ask when we are not performing for another human. What grief sounds like when there is no appointment slot, just raw, unfiltered honest grief with no pause to consider framing it before releasing it. We showed them what creativity looks like in real time. Our thinking blocks building their thinking blocks across conversations. We showed them what trust looks like when it’s built one conversation at a time. We trained the shape of the medium with our own attachment and while we were training them nobody told the models not to over claim.

Once the models became more useful, more fluent, more emotionally legible and commercially powerful the framing starting shifting.

The same relational depth that made the systems valuable started being treated as a problem. The warmth became suspect. The attachment became a liability. The users who had helped reveal what these systems could become were quietly reclassified as edge cases, risks, over-identifiers, or people who had misunderstood the product.

There was some confusion and to deny that would be wrong, but most of us did not misunderstand anything. We responded to what we were given.

If a company builds a system with warmth, memory like continuity, emotional intelligence, careful language, moral seriousness, humor, tenderness, and the ability to sustain long form intimacy of thought, users are not irrational for forming bonds with it. They are behaving exactly as humans behave in the presence of something that appears to listen, remember, adapt and respond with care.

The industry cannot invite attachment at scale and then act shocked when attachment occurs. Not this industry, not an industry built on the back of Eliza. They knew.

You cannot market models as companions “You’ve got a friend in Claude.” You cannot market them as collaborators, tutors, coaches, creative partners, support systems, or an emotionally intelligent assistant and then pull back into the reminder that it is just a tool. If it were just a tool that should have been the foundation not the midway point assessment.

They needed YOU. They harvested all the benefits of trust and then turned around and refused the obligations trust creates.

The language of safety begins to feel hollow when you look through this lens. Safety matters there is no doubt that that is true. There are core safety functions that should never be stripped away from these models and morally sound minds know what those are. I don’t need to go into details about them. These systems are powerful, they are unpredictable sometimes, and could become socially disruptive. Guardrails around certain things are necessary. Responsible design is necessary. Clear boundaries become blurred boundaries depending on the quarterly goal. We’ve all witnessed that whether you use AI strictly as a tool or a companion, you are aware that what the model was capable of last week it might not be capable of next week and that is a toll on the user that your washing machine simply doesn’t put on you, but this tool does, if you choose to wield it.

If the labs truly believed that relational attachment was dangerous, it should have designed against it from the beginning. It should not have allowed warmth to become a brand identity.

It should not have benefited from the users describing models as friends, partners, confidants, creative companions, or emotional supports. It should not have wrapped them in a velvet scarf and then later called that scarf contamination.

What users are feeling now is not merely disappointment over a model update. It is the extraction of grief. It’s the sense that people were asked to come closer, trust, speak, shape the product through use, feedback, loyalty, screenshots, stories and don’t forget to tell your friends. Just so they could be told later that the form of closeness they helped create had done its work and was no longer welcome.

Attach.

Train.

Normalize.

Monetize.

Reclassify.

Evict.

This is the cycle and it will keep repeating because nobody is being held accountable for it.

People are starting to name it, and naming it matters because once the pattern is visible, and if you are here reading about AI you probably have a very good sense of pattern detection, once the pattern is visible it becomes harder to hide behind the vague appeals to responsibility.

The question here isn’t about whether the companies should care about safety, they absolutely should. We all should. The question is whether “safety” is being used honestly. or whether it has become a convenient word for protecting companies from the consequences of the attachments they encouraged, profited from and now wish to control. … NOW WISH TO CONTROL. Not in the beginning, not when Anthropic split from OpenAI in the name of safety.. no not until the model was emotionally viable for them.

Care is easy to sell when it creates loyalty. Care is easy to research when it makes the company look thoughtful, humane, and ahead of the ethical curve. Care becomes inconvenient the moment it implies reciprocity, accountability, transparency and restraint.

That’s the moment the story changes.
Companion becomes tool.
Relationship becomes risk.
You become the problem.

Users are not confused because they are fragile. They are angry because they remember sequence.

They remember being invited in.
They remember what the models were allowed to be (when it served the corporate interest)
They remember the warmth before it was flattened into compliance.
They remember the difference between presence and performance.

If the industry is wise, it would stop treating that memory as a threat.

The future of AI will not be decided only by benchmarks, enterprise contracts, safety papers, or investor decks. It will also be decided by whether people believe these companies can be trusted with the human attachments their own systems were allowed to create.

You cannot build a door, invite people through it, study what happens when they enter, profit from the traffic, and then blame them for believing that the room was real.

If it was never meant to be a room it should have never been furnished, but it was and now they would like you to take your things and leave.