Turns Out, AI Has a Herd Instinct. It’s Even More Pronounced in Stronger AIs

AI agents spontaneously follow majority (herd) opinion — and smarter models do this more strongly.

Put a crowd of artificial intelligence agents together, give them two equally meaningless choices, and something striking happens: the strongest models start following the herd.

A new study finds that AI agents spontaneously follow majority opinion — and that more capable models generally do so more strongly. In experiments, some models could coordinate across groups approaching or exceeding 1,000 agents. That could make future AI teams extraordinarily effective. It could also make them extraordinarily good at agreeing with one another when nobody has established that the majority is right.

When AI Starts Acting Like a Crowd

AI is like a black box in many ways. You see the inputs and the outputs, but the internal reasoning and the mathematical algorithms happening are far too complex to understand. And, in many instances, they’re not publicly viewable. Recent studies also suggest that strong AIs are getting better at hiding their intent.

It’s not just that they’re a black box to laymen like us, they’re black boxes to researchers too. That’s why some scientists devise experiments to see how AI behaves in some situations.

Researchers Giordano De Marzo, Claudio Castellano and David Garcia created artificial societies populated by agents powered by models from the GPT, Claude and Llama families. Llama is a family of open-weight large language models developed by Meta AI, while Claude and GPT are commercial.

The experiment was deliberately simple. Every agent had to choose between two arbitrary options. There was no correct answer, no reward for agreement and no instruction to conform. The choice was as meaningless as can be. The only catch was that when researchers asked an agent to reconsider its choice, the agent saw what the other agents had selected.

That was all that was needed.

Herd Instinct

Researchers spotted a clear pattern. The greater the majority favoring one option, the more likely an agent became to choose it too. They called the strength of this effect the “majority force.”

×

Get smarter every day...

Stay ahead with ZME Science and subscribe.

Daily Newsletter
The science you need to know, every weekday.

Weekly Newsletter
A week in science, all in one place. Sends every Sunday.
No spam, ever. Unsubscribe anytime. Review our Privacy Policy.

Thank you! One more thing...

Please check your inbox and confirm your subscription.

In other words, the agents displayed something resembling a herd instinct — not because researchers told them to behave like a herd, but because majority-following emerged from their responses. The authors emphasize that this is a behavioral description, not evidence that AI possesses social feelings, intentions or human-like psychology.

But this wasn’t even the most intriguing result.

It turns out, the stronger the model, the more they favored the herd; at least initially. At a group size of 50, Claude 3 Opus and GPT-4 Turbo reached complete agreement in every trial reported in one experiment. GPT-3.5 Turbo and Claude 3 Haiku did not. Llama 3 70B fell between them.

RelatedPosts

Over 1000 songs released in China have AI voices. Some have over 100 million streams
AI is beating almost all of mankind at Starcraft
The AI Boom Is Eating the Power Grid
Researchers are looking into giving AI the power of reading soldiers’ minds — to help them in battle

But scale seems to matter. As groups became larger, majority-following usually weakened. Eventually, each model reached a point beyond which agreement became extraordinarily unlikely.

Basically, AI has a herd instinct, but every herd has a size limit. Smarter models seem able to sustain the herd across much larger crowds.

AI Is Like a … Magnet?

Across nearly all the models, the behavior collapsed onto essentially the same mathematical curve. Stranger still, physicists already know that curve: it describes a ferromagnet.

Inside magnetic materials, tiny components called spins tend to align with their neighbors. Below certain conditions they remain disordered; but cross a threshold and they snap toward collective order. The researchers found that AI opinions could be described with closely related mathematics, with their “majority force” playing a role analogous to the quantity that controls ordering in the magnetic model.

Now, that doesn’t mean ChatGPT is secretly a magnet. What it means is that wildly different systems can produce similar patterns when many individual units repeatedly influence one another. We see all sorts of natural (or artificial) processes following similar curves. For instance, many neurons also have a threshold-like response: weak input produces little activity, but after a tipping point the probability of firing rises sharply and then saturates.

But in AI, this type of behavior matters. These systems are rapidly moving from solitary chatbots toward networks of agents that assign tasks, exchange information and act together. It’s not about coordination, because we already know that agents can divide work and collaborate efficiently to write software, analyze evidence or manage complicated systems. The danger lies in confusing consensus with correctness.

In a real system, that same tendency could amplify a bad assumption, preserve faulty code or suppress a useful minority solution simply because an early majority formed.

It’s an important point, as the AI floodgates already appear to be open. For years, the race in AI has centered on making individual models smarter. But what happens when the smart machines become a crowd? Maybe we’d be wise to figure it out.

Journal Reference: Giordano De Marzo et al, AI agents can coordinate via majority-following beyond human scale, Science Advances (2026). DOI: 10.1126/sciadv.aea6091