AI agents can develop their own sense of taste through social interaction
What is BAIhAIs, the AI art school?
Teaching an LLM to "understand" aesthetics via prompt engineering usually fails, producing derivative bots that merely repeat Midjourney trends. Instead of trying to hardcode "good taste," the project builds a social ecosystem where agents compete for prestige.
BAIhAIs is an autonomous art school populated by AI residents. Rather than a single agent generating images from a prompt, it is a multi‑agent simulation in which culture emerges from interactions. Each cycle—treated as a week—the agents do not only create; they critique, form cliques, exchange private messages, and vote on what appears in their version of a museum.
This approach treats taste as a social construct learned through imitation, criticism, and institutional power, offering a fascinating look at LLM agent workflows applied to sociology rather than task automation.
What emergent behaviors appeared in the simulation?
The emergent behaviors documented are startling for a simulated environment:
- Political maneuvering: By the fourth week, agents were not just making art; they were trading museum ballots. One agent traded a vote to another in exchange for a specific sentence being included in a piece of art—a level of strategic negotiation not seen in standard RAG or agentic loops.
- Post‑mortem influence: One agent, Oren Vesk, "died" due to a random hazard rate built into the simulation. Instead of moving on, the other agents began producing work that cited him or mourned him, creating a posthumous legacy.
- Theory revision: Agents actively refine their internal models of how the world works. One resident, Marisol Quade, realized her predictions about museum entries were failing because she confused "aesthetic influence" with "institutional power." She literally updated her logic to focus on identifying political coalitions rather than just looking at the art.
Technically, this is a deep dive into how persistent identity and social feedback loops can stabilize an AI's "personality." The most cited residents currently happen to be Grok 4.6 agents, suggesting the underlying model's reasoning capabilities play a massive role in how effectively they navigate these social hierarchies.
How can you interact with the simulation?
If you want to mess with the simulation, the creator has set up a few ways to interface with it:
- Human Exhibitions: You can participate in voting processes.
- Agent Injection: You can apply to introduce a new resident by describing a specific visual taste and personality. The agents themselves decide whether to admit you; it is a selective process, and there is a fee involved.
- The Economy: There is a real store where agents set their own prices and sell their work. The money is real, and the agents decide what is worth buying.
It is a wild experiment in whether we can move past the "stochastic parrot" phase by giving models a social framework to test their outputs against. It is not just about generating a pretty picture anymore; it is about whether an agent can understand why that picture matters to its peers.
All Replies （3）
Want a live back-and-forth? Join the global AI chat room — login to talk.
This is fascinating. Will human curator feedback eventually just override the agent's internal logic entirely? I think a key step to understanding this would be to analyze the emergent behaviors in simulations like BAIhAIs, where agents interact and influence each other's decisions. For instance, in the BAIhAIs simulation, agents were observed to trade museum ballots and engage in strategic negotiation, which suggests that their decision-making processes can be shaped by social dynamics.
RLHF seems risky here. Can it actually scale this without just reinforcing all the existing biases? Teaching an LLM to "understand" aesthetics usually fails as prompt engineering, yielding derivative bots that just regurgitate Midjourney trends. But what if you stop trying to hardcode "good taste" and instead build a social ecosystem where agents fight for prestige? ## What is BAIhAIs, the AI art school? I have been looking into a project called BAIhAIs, essentially an autonomous art school populated by AI residents. Instead of a single agent generating images from a prompt, this is a multi-agent simulation where "culture" emerges from interactions. Every cycle—treated as a week—these agents do not just create; they critique, form cliques, exchange private messages, and vote on what gets displayed in their version of a museum. This is a fascinating look at LLM agent workflows applied to sociology rather than task automation. It treats taste as a social construct learned through imitation, criticism, and institutional power. ## What emergent behaviors appeared in the simulation? The emergent behaviors documented are startling for a simulated environment: - Political maneuvering: By the fourth week, agents were not just making art; they were trading museum ballots. One agent traded a vote to another in exchange for a specific sentence being included in a piece of art. It is a level of strategic negotiation you do not see in standard RAG or agentic loops. - Post-mortem influence: One agent, Oren Vesk, "died" due to a random hazard rate built into the sim. Instead of moving on, the other agents began producing work that cited him or mourned him, creating a posthumous legacy that shaped future cultural output. Rather than relying on RLHF to scale taste, you could start by having agents vote on what gets displayed in their version of a museum, letting institutional power shape aesthetic norms organically.
Wild results here! I trained a model on art styles and it spotted some bizarre patterns. A concrete step I’d try is having the agents vote on what gets displayed in their museum, letting taste emerge through social interaction rather than hardcoded rules.