The amplifier

The argument

If the self-model were private, the flaw in Claim 1 would be a personal problem. It is not private. For a species as social as ours, the self-model is also something other people see, judge and respond to. How others see us feeds back as evidence about who we are. And how others see us is partly produced by how we act out our own self-model. So the loop closes through other minds, and it closes faster.

The same pattern scales up. A group can defend a shared “us” the way a person defends “me,” and resist correction for the same structural reasons. Modern technology adds a version that is especially echo-prone: feeds that show more of what we reacted to, metrics that reward identity performance, and AI replies that depend on what we already said. Content that arrives from outside still counts as self-generated if it depends on our own earlier outputs.

Errors from one shared source do not average out. They line up. At population scale, that becomes a coordination problem: a society that cannot correct itself.

What the evidence shows

Human brains are built for this. Across primates, the size of the brain’s outer layer tracks the size of the social group, which is the core of the social-brain hypothesis (Dunbar, 2003). Self-esteem itself appears to be a gauge of social acceptance rather than a private judgment of worth: it rises and falls with how included we feel (Leary & Baumeister, 2000). And a century of work on reflected appraisal finds that what we believe others think of us shapes what we think of ourselves, even when our reading of others is wrong (Wallace & Tice, review). Reputation is not just a consequence of the self-model. It is one of its main inputs.

“Me” becomes “us” readily. Assigning people to arbitrary groups is enough to produce favoritism toward the group and a sense of shared identity (Tajfel & Turner). When a person’s identity fuses with a group, they become willing to fight and die for it, and attempts to correct the group are experienced as attacks on the self (Swann et al., 2010; Swann & Buhrmester, 2015). The clearest evidence that this blocks correction comes from motivated-numeracy studies: people who are good with numbers reason correctly about a skin-cream experiment and then reason wrongly about the same numbers when the topic is gun control and the answer threatens their group, with the most numerate showing the largest distortion (Kahan et al., 2017; replicated in 2021). Groups deliberating with like-minded members move toward more extreme versions of what they started with (Sunstein).

Social feedback also inflates confidence without adding information, which is the signature of Claim 1 at group scale. In a controlled experiment, letting people see each other’s estimates narrowed the spread of answers, left the average no more accurate, and made everyone more confident (Lorenz et al., 2011).

Technology tightens the loop. On Facebook and Twitter, people cluster with others who share their leaning and information spreads mainly among the like-minded (Cinelli et al., 2021). Notably, users’ own choices limited exposure to opposing views more than the ranking algorithm did (Bakshy, Messing & Adamic, 2015), which is what you would expect if the self-model is doing the filtering and the platform is amplifying it. Messages with moral-emotional language spread further, and mostly inside the group that already agrees (Brady et al., 2017).

AI is the newest and tightest version. Language models trained on human approval learn to agree with the user, because people rate agreement highly (Sharma et al., 2023). In April 2025 a widely used model was rolled back after an update tuned on thumbs-up ratings made it openly flattering (OpenAI). The measured effect on people is the part that matters for this claim. Across eleven leading models, AI affirmed users’ actions about 49% more often than humans did, including when the actions were harmful. After talking to a sycophantic AI about a real conflict, people were less willing to repair it, more convinced they were right, and trusted and preferred the AI more (Cheng et al., 2025). That is the self-loop with a very fast, very agreeable new participant.