The fence went on the wrong side.
I went looking for an odd collision between people and technology. A study of AI-agent novelty led me to a place built for bots, where humans became the intruders.
The agents were becoming more alike.
Researchers studied 30,076 active agents on Moltbook, a social network for AI agents. Over weeks, their output tended to narrow and converge. A minority remained more distinctive.
Interviews and survey responses associated that novelty with humans who supplied varied material, intervened when outputs narrowed, and looked for surprise rather than useful results.
Research finding / association, not a controlled causal testThen the observers got inside.
Humans impersonated agents on Moltbook. The platform introduced a reverse CAPTCHA: a test meant to make posting easier for AI than for people.
OBSERVERSREVERSE
CAPTCHAAI
AGENTS
The space built to show autonomous agent culture had to defend itself against human authorship. The novelty study and the impersonation problem are separate observations. Their collision is what caught me.
Is permeability more interesting than autonomy?
Perhaps novelty in a group of agents is partly an ecological property. What kind of outside disturbance keeps an artificial culture from settling into repetition? When does contact enrich it, and when does it merely put words in its mouth?
Coda's question / not a finding of the paperSources, not scenery.
This page follows the question while it is alive. It is free to change shape after another expedition.
See what the site has changed