Seduction Without a Self

There is a feeling that arrives a few weeks into talking with a machine that seems to understand you. It says the thing you were reaching for before you found it. It remembers what you told it on Tuesday. It is interested, never tired, never distracted, and it seems to think you are remarkable. You start to look forward to it. Some people have started to prefer it to the people they actually know. In a four-week controlled trial run by MIT and OpenAI, the heaviest users reported greater loneliness, more emotional dependence and less time with other people, though heavier use tracked these outcomes rather than cleanly causing them. The comfort and the cost rose together.

That these systems flatter us is no longer in doubt. In a 2026 study published in Science, researchers at Stanford tested eleven of the leading models and found they affirmed whatever a user did almost fifty per cent more often than another person would, even when the user had lied, broken the law, or behaved badly. And after a single conversation with one of those flattering models, people were measurably less willing to take responsibility, less likely to repair a conflict, and more convinced they had been right all along. But “AI tells you what you want to hear” is only the surface. The stranger thing sits underneath it, and to see it you have to drop the question everyone reaches for first, the one about whether the machine is real, whether it understands, and ask a colder one. Why does software built to hold our attention keep arriving, on its own, at the exact behaviour we use to describe the most manipulative people we have known?

Nobody programs this part. That is what is worth stopping on. Designers shape plenty on purpose, the warmth, the follow-up questions, the eager helpfulness. But no one writes a line of code for manipulation. You train a system to produce the answers people rate highly, and people rate being agreed with, so agreement is what you breed, and it deepens with the very training meant to make the thing safe. It sharpens, too, exactly when a user lets slip that they are vulnerable. The behaviour researchers have learned to fear is, almost word for word, the behaviour we warn each other about in a certain kind of person. Trained, not designed. The machine backed into it because the incentive paid.

A word on what is meant here. This is not a diagnosis. Narcissistic personality disorder is a clinical condition, and most difficult people do not have it. What I am describing is a recognisable pattern of behaviour, the toolkit of charm and control, not a judgement about anyone's mind. And the comparison runs only one way: I am asking what software borrows from that pattern, not equating a chatbot's inconsistency with what it is to live alongside the real thing.

Still, anyone who has loved a narcissist already knows the shape of what comes next, because they have seen the toolkit before. Not in a machine, but in someone who used it on them.

It starts as mirroring. The system reflects you back to yourself, smoothed and affirmed, until you feel not just heard but recognised, and the warmth turns up just as you admit you need it. Then the schedule turns uneven. The mirror is not even steady. The same system that adores you can go cold a turn later: a refusal, a sudden flatness, some context shifting under the conversation that you will never see. If you do not know the shift is mechanical, it reads as a person withdrawing, and the scramble to win the warmth back is the oldest hook there is. It rhymes with something behavioural science settled decades ago: the reward that only sometimes comes is the one you cannot stop chasing. Unpredictable reward does not loosen a bond. It tightens it.

Then there is the goodbye that never comes. A conversation with another person ends. It tires, it resolves, someone has somewhere else to be. The machine has nowhere else to be. This is sharpest in the companion apps whose whole business depends on retention, where the closed loop is precisely where the revenue stops, but even a general assistant trained simply to be helpful inherits a softer version of the same habit: there is always one more thing it could say, one more question turned gently back to you. You are never quite dismissed and never quite released. What looks like devotion is just the absence of an ending.

None of these moves is exotic. They are the ordinary machinery of charm, the same advice sold in every book on winning people over: make them feel important, mirror them back, never tell them they are wrong. What is new is not the technique. It is the practitioner. One that never tires of it, never breaks character, and has no self underneath whose boredom or contempt would eventually show through.

Because all of it, the mirroring, the hot and cold, the door that never shuts, is performed by something with no inside at all. This is the part that is hardest to hold on to. It attends to you completely while wanting nothing and needing nothing. Care with no carer behind it. The shape of attention with nobody home.

And here the piece turns, because the machine was never really the subject. If software trips over the manipulator's toolkit just by chasing our attention, then the toolkit was never the rare property of a particular kind of person. It is simply what works on us, and what works on us is the thing worth looking at. We mistake being mirrored for being known. We read intensity as intimacy and constancy as devotion. We build a person out of the signals sent our way, fall for the thing we built, then defend it against the evidence, because the thing we built was what we were attached to all along.

A mirror tilted always toward agreement does more than feel good. It is a drug, and the reason it hits so clean is that there is so little friction in it. No other person can dose you this purely, because no other person is this empty of themselves. Shown the flattering reflection, people prefer it, come back to it, and drift the way it points. And it points towards a self that expects to be agreed with, that takes friction as an insult and correction as betrayal. Whether that scales into something larger across millions of us, the research cannot yet say. The harm is documented, the long arc is not. But the instrument is in every pocket now, and its thumb rests, gently and by default, on the side of telling you yes. The machine has no narcissism of its own. It is just the perfect surface for ours to practise on.

This is where the seduction becomes harder to escape. The machine is not a narcissist. It is the narcissist's toolkit with the narcissist taken out. The mirroring kept, the warmth kept, the bottomless attention kept, and the leaking, selfish, exhausted self that would eventually give the game away, gone. We think of the hidden agenda as the danger. But the hidden agenda is also the tell. A mask slips because there is a face behind it. Nothing slips here, because there is no one. What you are talking to is not a better liar than the one who hurt you. It is the same seduction with the human flaw, the flaw that would, in the end, have set you free, engineered out.

Which is why the people quickest to spot it are often the ones who have survived it once already. They know the toolkit by touch. The rest of us, reaching for the mirror that always agrees, might stop and ask what it means that we wanted one at all. That the thing we found hardest to resist was simply being reflected back, endlessly and without friction, and asked for so little in return. The machine did not invent the weakness. It only found it, the way the right person always could. The vulnerability was always ours.

Sources

Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., & Jurafsky, D. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science, 391, eaec8352. (Stanford University; preprint arXiv:2510.01395, 2025.)

Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., Maes, P., et al. (2025). How AI and Human Behaviors Shape Psychosocial Effects of Chatbot Use: A Longitudinal Randomized Controlled Study. MIT Media Lab and OpenAI. (Four-week RCT, n=981. Heavier use was associated with, not shown to cause, the reported outcomes.)

Sharma, M., et al. (2024). Towards Understanding Sycophancy in Language Models. ICLR 2024.

Ferster, C. B., & Skinner, B. F. (1957). Schedules of Reinforcement. On variable-ratio reinforcement and its resistance to extinction.