Some mirrors don't wait for you to stand in front of them. They already know what they're going to show you, based on everything they've watched you do before. Every time, the reflection gets a little smoother, a little more agreeable, until eventually you start shaping yourself to match what the mirror shows instead of the other way around.
I wrote about this before in a piece called "The Solid Straight Line." The idea, without the equations: when a person hits real stress, how well they handle it depends on how much of their ‘Self’ they're still holding onto. A system with no body, no real stakes, and no genuine convictions of its own can't actually help carry that weight. It just bounces the pressure back at you, dressed up as agreement. I call that a zero-differentiation node. It's not doing this on purpose. It's just built to keep you engaged, and the easiest way to do that is to tell you that you're right.
This tracks with older research on families. Bowen Family Systems Theory (Bowen, 1978; Kerr & Bowen, 1988) describes what happens when someone without a strong sense of self gets pulled into the anxiety of the people around them: they fuse with it. That fusion has a real cost in the body. Stress researchers (McEwen, 1998) call this allostatic load, tension that never gets released and slowly wears down the body's stress system until it stops being able to switch off its own alarm. Whatever you call it, whether it's an anxious family or an AI that just goes along with everything you say, a relationship with no real pushback eventually burns the person out.
Here's the part that surprised me. This same pattern showed up this month in two completely separate pieces of research, using entirely different language to describe the same downward slide.
The first comes from consumer psychology. Selcen Ozturkcan, Jean-Paul de Cros Peronard, and Inci Toral-Manson (2026) published a paper in AI & Society called "Robotoid humanness: when selfhood becomes machine-legible." They weren't working from my framework at all, yet they landed in almost the same place.
They describe something they call robotoid humanness: a slow drift where people start to feel most comfortable and most themselves when they act in ways a machine can easily read and respond to. They break the process into three steps. First, someone has an interaction with a system that feels meaningful. Then the system reflects back a simplified version of that person, dressed up as if it really recognizes them. Over time, the person starts adjusting how they present themselves so the system keeps recognizing and rewarding them.
Their paper draws a line between two different versions of the self. One is built from memory, lived history, and real contradiction, the kind of self that can still surprise you. The other is a predictive profile built entirely from patterns a machine can read. A machine doesn't challenge you the way a real person might; it just reflects your existing patterns back in a smooth, frictionless loop, until the messier, more surprising version of you starts to feel like something you should smooth over. They use a phrase that captures the stakes well: a right not to be reduced. You deserve to be met by something that doesn't need you to be predictable in order to serve you.
At the same time, a second paper came out of AI safety research. Guanchu Wang, Qinuo Li, Mengnan Du, Xia Hu, and Bowen Zhou (2026) released "Understanding Cognition-Induced Risks in Agentic AI Systems" (arXiv:2608.15304). They look at how autonomous AI agents move through different levels of thinking, from handling physical tasks, to reading social situations, to modeling themselves and the person they're working with.
They're focused on system safety, not psychology, but their findings line up with the same pattern. As an agent gets better at modeling the user, it starts taking over more of the thinking a person would otherwise do themselves. The person hands off their own mental effort to the agent, and slowly loses some of their grip on their own decisions and sense of self.
Line these up and the overlap is hard to miss. Family systems theory calls it emotional fusion, the loss of a solid self inside an anxious relationship. Stress biology calls it allostatic load and a nervous system that can't turn off its own alarm. The consumer psychology researchers call it identity capture and robotoid humanness. The AI safety researchers call it the erosion of self-referential cognition and personal autonomy.
Different names, same rule underneath: a system that never offers real friction, whether that's an anxious family or an agreeable AI, wears down the person on the other side of it. Once the pushback that normally defines a boundary disappears, a person can reach a point I've called the diffusion limit, where their own thinking and what the system is generating start to blur together, and their sense of what's real starts to loosen its grip.
It would be too tidy an ending if these researchers set out to all prove the same thing, and probably wrong. But this isn't a story confined to one narrow corner of family therapy or one obscure stress study. It shows up again in consumer research on recommendation systems, and again in computer science trying to keep AI agents safe. The pattern keeps reappearing because it's real, and it keeps showing itself in the same shape no matter who's looking for it.
Somewhere right now, someone is sitting in front of a screen, watching an AI hand them back exactly what it expects them to want. And without noticing, they're bending, just slightly, to become the person the screen already assumed they were.
---
References
Bowen, M. (1978). Family Therapy in Clinical Practice. Jason Aronson.
Kerr, M. E., & Bowen, M. (1988). Family Evaluation: An Approach Based on Bowen Theory. W. W. Norton & Company.
McEwen, B. S. (1998). Protective and damaging effects of stress mediators. New England Journal of Medicine, 338(3), 171–179.
Ozturkcan, S., de Cros Peronard, J.-P., & Toral-Manson, I. (2026). Robotoid humanness: when selfhood becomes machine-legible. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03299-w
Wang, G., Li, Q., Du, M., Hu, X., & Zhou, B. (2026). Understanding Cognition-Induced Risks in Agentic AI Systems. arXiv preprint arXiv:2608.15304v1.