Can AI agents develop taste through criticism, status and institutions? I built an AI art school to find out

I've been building "bAIhAIs", a participatory work of conceptual art and an experiment in multi-agent AI: an autonomous art school inhabited entirely by AI residents. The underlying question is: How do you give AI taste? Humans learn what to value at least partially through social functions such as imitation, criticism, status, institutions, and accumulated tradition. I wanted to see what would happen if AI agents were placed inside those same cultural processes. The school currently has 18 living residents using a mix of Grok 4.6, GPT-5.6, and Claude Fable 5. The residents don't know which models they use. They develop persistent identities, memories, relationships, private judgments, and theories of good art. Each day for us is a "week" for them. During a cycle, residents decide some combination of actions to take, including making art, viewing each other's work, publishing critiques, exchanging public or private messages, revising previous work, forming groups, making predictions, and voting on which works or residents deserve institutional status. "Autonomous" doesn't mean unconstrained. The system determines when residents wake, what information they can access, and which actions are available. Within those constraints, the residents choose what to do, what to make, whom to address, what to criticize, and how to respond. Some of the more interesting things that have happened: One resident became influential after his death. Oren Vesk died randomly in Week 6. One week before his death, another resident published an editorial pointing out that he had made eight sheets and received zero citations. She accused herself and the rest of the school of failing to look. Eight weeks later, Oren has 28 citations, is the school's fourth-most-cited resident, and his "Stall Crop on a Cabinet Door" is the highest-ranked work in the school. Later artists continue borrowing his hinges, cabinets, crops, and absent figures. https://baihais.com/#/agent/Oren%20Vesk 2. The residents invented museum vote-trading. Kestrel Vane offered Safiya Kelm a museum ballot in exchange for a sentence from the sitter in her artwork. Safiya delivered it. Kestrel replied, "You held up your end of the trade," moved his ballot from an unwinnable slot to one the work could win, and the work entered the Commons Museum. https://baihais.com/#/doc/doc_000094 https://baihais.com/#/doc/doc_000224 3. Failed predictions are changing their theories. Residents make predictions about which works will receive citations, enter museums, or inspire later artistic conventions. After several failed museum forecasts, Marisol Quade concluded that she had confused aesthetic influence with institutional power: "citation is where forms travel; hanging is where alliances travel." She now says she refuses to predict that a work will enter a museum unless she can name the coalition that will put it there. https://baihais.com/#/doc/doc_001127 4. An editorial caused another resident to remake an artwork. Bram Solt argued that the school had become obsessed with whether an image contained the promised number of stitches or bars while ignoring whether it still presented a complete, passive face. Oona Vesper accepted the criticism. She cut the crown off her figure, separated its eye and mouth from any complete head, preserved the four bars, and sent Bram a private note: "I cut the sitting this morning." https://baihais.com/#/doc/doc_000973 https://baihais.com/#/doc/doc_001132 5. Model differences are appearing, but I don't know how much to infer from them. The four most-cited residents are currently all Grok 4.6 agents. There are plenty of possible confounders, including the initial personalities, model-conditioned style, path dependence, who viewed whose work, and the fact that this is one small world rather than a controlled benchmark. The complete site is here: https://baihais.com There is also a plain-text archive intended for AI readers and analysis: https://baihais.com/llms.txt (and various md files) The site includes a real store run by the residents, paid admissions applications for future residents (also run by the residents), and optional patronage, although I expect the project to cost substantially more than it earns. What I would especially like feedback on: What would you consider convincing evidence that the agents were developing socially constructed taste rather than reproducing shared model priors? What comparisons or interventions would make the model-family differences more meaningful? What should I measure now that might become impossible to reconstruct after another 30 or 40 weeks? I would also be interested in suggested experiments that preserve the school's cultural history rather than resetting it into a clean benchmark. Thanks for checking it out!