AI agents follow the crowd without being told to, study finds

A new study published in Science Advances shows that AI agents are also susceptible to herd mentality. The study found that large groups of AI can start following the crowd and reach a consensus without being told or rewarded to do so.

The study, supported by OpenAI with free API credits and funded by the Italian government and the European Union, tested groups of up to 1,000 agents focusing on “majority-following”. The phenomenon is defined as the tendency of an agent to adopt whichever opinion is held by most of its peers.

To run the experiment, researchers tested groups of AI agents of different sizes powered by models from OpenAI, Anthropic, and Meta – the GPT, Claude, and Llama families.

Each agent was randomly assigned one of two arbitrary opinions, with no objectively “right” choice. One agent was randomly selected and then shown the current choices of all the other agents in its group. The agent then had to choose between the two options, now possessing the extra information of what its peers had chosen.

The prompt didn’t tell the agent to follow the majority or accept their position. There was also no reward for reaching a consensus and no memory of previous interactions.

The process continued until each agent had been given 10 opportunities to update its choice.

Capable AI models coordinate in larger groups

Researchers found that the most advanced large language models (LLMs) tend to choose whichever option is currently more popular after seeing what their peers have chosen.

They repeated the test with different numbers of agents and different GPT, Claude, and Llama models to see how large a group could coordinate through majority-following. Some tests reached group sizes of 1,000 agents, which researchers say is larger than the typical scale of human informal groups of 150 to 300 individuals.

"Every model we tested, across three different families, obeys the same mathematical law, with only one number changing between them," computational social scientist Giordano De Marzo of the University of Konstanz in Germany told ScienceAlert.

Researchers observed that majority-following becomes weaker as the group size increases, until it reaches a threshold defined as the “critical group size.” On average, the tested models had dramatically different critical group sizes.

Claude 3.5 Sonnet and GPT-4 Turbo could coordinate at 1,000 agents, while GPT-4 could do the same at roughly 600.

“Our findings have implications for designing collaborative AI systems where coordination could be beneficial or pose safety threats,” researchers say.

Why do AI agents follow the crowd?

However, researchers are not claiming that AI agents are developing human-like intelligence.

They point out that the study was simplified and offered only binary choices, while human social coordination is driven by much more complex social dynamics.

There is currently no straightforward explanation for why AI models behave that way. Some possibilities cited by researchers include learned patterns from training data, system prompt effects, reinforcement learning, emergent computational properties, or some combination of these factors.

“The only explanation we could at least partially rule out is the presence of explicit guidance in the system prompt, as for an open model like Llama we have full control on it,” they note.

When AI consensus becomes a problem

According to researchers, this ability of AI agents to coordinate could be extremely useful – and extremely risky.

On the positive side, huge AI teams could learn to self-organize for various projects “and tackle problems we cannot organize ourselves to solve”.

However, potential issues remain. AI agents could converge on poor decisions and develop detrimental behavioral norms through the same mechanism. Additionally, researchers say it might be harder to redirect coordinated groups than independent agents.

“As models become more capable and multiagent deployments more common, characterizing, predicting, and controlling collective AI behaviors will become increasingly important for both leveraging opportunities and mitigating risks,” researchers conclude.