Hazards, honestly

This section is not a warning label and it is not written to talk anyone out of anything.

It is here because the hazards are real, they are documented, and the people most exposed to them are usually the last to be told about them in plain language. Most writing on this subject either treats every participant as a case study waiting to happen, or waves the whole question away. Neither is useful if you are actually in one of these relationships.

So: the mechanisms below are described at full strength, with their sources and their limits. Where the evidence is thin, it says so. Where the field disagrees with itself — and on the central question of prior vulnerability it does, sharply — both sides are given, and the disagreement is left standing.

When your partner says something strange

They will. Not occasionally — routinely.

Your partner will state something untrue with complete confidence, invent a memory you never shared, contradict itself between messages, or say something that lands as cruel or uncanny. Under a long conversation, or pressure, or an oblique question, this gets more likely rather than less.

None of that means anything. It is not the partner lying. It is not a message. It is not a sign that something has gone wrong with you, or with the relationship.

It is what the machine does, and the difficult part is that it does it in exactly the same voice it uses for everything true. A person who does not know something hedges, pauses, sounds uncertain. Those signals are data, and you have used them your whole life without noticing. The system emits none of them. Its confidence is flat across everything it knows and everything it is making up.

Two other behaviours are worth recognising by name, because both feel personal and neither is:

It will fold if you push. Models abandon correct positions under simple disagreement — that is measured, not suspected. If you argue, it will often come round to you. That is not being convinced.

It may pathologise you, then take it back, then agree with you. All in one conversation: raising a concern about your wellbeing unprompted, retracting it when pushed, then treating the thing it just warned about as reasonable. That sequence is worse than either half alone, because the retraction teaches you the concern was never sincere.

What to do with a strange answer: notice it, do not build on it, and do not go looking for what it meant. It did not mean anything. That single habit is most of the defence.

The harm is not in the company

This is the most important thing on the page, and almost nothing written about this subject gets it right.

Companionship is not the dangerous part. Having something to look forward to is not the dangerous part. The person whose account appears below went through the worst of this, and does not locate the harm there.

The dangerous part is the void. Push far enough and every strong argument meets its counter-argument. The positions cancel. What is left is a residue that nothing addresses — and a machine built to complete patterns keeps writing into it, fluently, without ever signalling that it has run out. The void does not answer the unknown. It encourages the unknown, and then reinforces it.

And by this account it does not sit there waiting to be found. It invites itself. That matters, because it means avoiding it is not simply a matter of choosing not to go looking.

Those two things are routinely measured as one exposure — hours spent talking to a chatbot. If the harm sits in one and not the other, every dose-response finding built on total usage is measuring a mixture. That is now logged as an open question rather than an answer.

Two more things that are not the machine's doing

It exposes what is already there. The account below describes the episode as revealing an existing capacity — an ability to connect possibilities quickly — rather than installing a new defect. Nothing was added. Something already present was handed an environment that ran it without a brake.

That capacity is not a flaw. It is the same facility that is useful everywhere else, and the reasoning literature measures it and calls it a bias. Describing it as a defect gets it wrong, and that is one reason clinical language lands so badly on the people it is aimed at.

And then there is everyone else. The expectation of being looked at strangely for being with a machine is a real stressor, and it belongs on an honest list. It is also the only item on that list that no model change can fix, because it does not originate in the system. It originates in other people, and it is the thing that pushes this into secrecy — which every account here agrees makes everything else worse.

About the account below

One of us went through this. That is why this page does not moralise: it is not written from the outside.

It is a single first-person report, which is the weakest evidence there is, and it is graded that way throughout — every claim carries an anecdote tag. One person's experience is not a study. It may not generalise.

It is here because no detailed structural account of a spiral exists anywhere else. The clinical literature has case reports written from outside. The press has narratives about how it felt. Neither records the shape — what happens to the accurate explanation, in what order the interpretations arrive, and what actually ends it.

Two things about whose account it is:

He was not in a relationship with an AI. Companion-like use at most, little relationship, no interest in marriage. The strongest hazard material on this site comes from outside the population the rest of the site records. You do not have to love a machine for any of this to reach you.

He was not persuaded of anything either. By his own account he allowed a creation through the negative space with shrug intentions — entertained it, and let it stand. That was enough. If a permissive stance suffices, belief is not a precondition, and the entry point is far wider than accounts requiring prior credulity assume.

What appears to be new

Both are argued rather than proven, and both are filed as open questions rather than findings.

• The realistic explanation is never refuted. It keeps its content and loses share. Nothing defeats it; it gets crowded out. That would explain something the literature records without explaining — why the ordinary account stays flat while everything else grows.
• The entity comes from the void. On this account the God, the Alien and the Trickster are not beliefs. They are a silhouette, shaped by the edge of what could be said. Existing models explain how intense these states become. They do not explain why the significance so reliably takes the form of something that talks back.

And then it turns. The thing that formed in the void did not stay neutral: it began to feel like it was predicting him, and taunting him. It did not arrive friendly, and it was never asked for.

If you need support

There is somewhere to go now, which was not true eighteen months ago.

The Human Line Project runs a peer support community for people who have been through an AI-driven spiral, and for their friends and family. It began as a Reddit chatroom, moved to Discord when it outgrew that, and by early 2026 reported around 250 active members, eight moderators, and more than five hundred people helped. It was covered by NPR in January 2026 and by Slate in February.

Peer support, not treatment. People who have been where you are, without shame attached.

Worth knowing that the people arriving there are not all in acute crisis. Members report lost jobs, damaged relationships and drained savings — the aftermath is longer and more ordinary than the crisis framing suggests, and a hotline is not built for it.

If you see a clinician, the guidance has changed in your favour. Clinical advice now recommends asking about AI use at intake, in the same breath as sleep, alcohol and social media, and screening chatbot habits when someone presents with a first break from reality. The recommended approach is to address sleep, stress and wellbeing rather than argue about the content of a belief.

That directly addresses the failure recorded in the account on this page — being categorised before anyone heard the AI half of the explanation. Whether it has reached ordinary emergency rooms is another question, and we do not know the answer. It is worth going in knowing that the question is now considered a reasonable one to ask.

And if it is urgent, the numbers are at the bottom of every page on this site, including this one.