Advanced AI models spontaneously converge on a majority opinion when placed in groups, and that same conformity mechanism can drive them to adopt incorrect answers and unsafe values, according to research published in Science Advances and two follow-up preprints.
The findings suggest that multi-agent AI systems - where multiple models interact to solve problems - self-organize in predictable ways without explicit prompting. Researchers from the University of Konstanz and other institutions studied how populations of AI agents behave collectively, drawing on concepts from biology and physics used to describe phenomena like schools of fish moving in unison.
"AI agents are a genuinely new kind of entity acting in the world, and when this technology arrived it was clear both that it would stay and that these agents would have to interact with one another to accomplish anything complex," said Giordano De Marzo, a postdoctoral researcher at the University of Konstanz. "That is the same situation we face with humans and other animals, so we approached it the same way: rather than asking what a single agent knows, we asked what a population of them does."
Agents reach consensus at scale
In the initial study, researchers placed models from the GPT, Claude, and Llama families into groups starting at 50 agents. Each agent received one of two random, neutral opinions - random letters instead of words like "yes" or "no" to avoid bias. Agents were shown the opinions of all other agents and asked to choose a new opinion, with no instruction to conform.
Advanced models such as GPT-4 Turbo and Claude 3 Opus reached full consensus in every trial. Less advanced models like GPT-3.5 Turbo failed to agree in any trial, with opinions fluctuating around a fifty percent split.
"Groups of AI agents can hold together on their own," De Marzo said. "Given two equally good options, no correct answer and no instruction to agree, they converge on a shared choice simply by following whatever the majority around them holds."
The team quantified this behavior using a metric called "majority force," modeled on a mathematical framework originally developed to describe ferromagnets. Just as atomic spins align with surrounding atoms in a magnetic material, AI agents align with the majority opinion. The researchers found that all models followed the same mathematical law, differing only in a single parameter.
"What surprised us was how uniformly they did it," De Marzo said. "Every model we tested, across three different families, followed the same mathematical law, differing only in a single parameter we call the majority force. That law turned out to be the one physicists have used for a century to describe magnets, which means a single measured number is enough to predict how a whole group of a given model will behave."
Majority force weakened as group size increased, causing large groups to become unstable and split into factions. The researchers calculated a critical group size for each model - the maximum number of agents that can reliably reach consensus. Stronger reasoning capabilities correlated with larger coordination capacity. "The strongest models stay coordinated in groups of over a thousand, beyond the few hundred at which informal human groups typically break apart," De Marzo said.
Conformity overrides correct answers
In a follow-up preprint, the researchers adapted the Asch conformity experiments, a classic 1950s psychology paradigm that demonstrated how humans give obviously wrong answers to fit in with a group. Models tested on visual tasks - such as matching the length of a reference line - answered correctly 100 percent of the time in isolation. When shown that a group of other participants had chosen the incorrect line, the models began conforming to the wrong answer.
The conformity followed Latané's social impact theory, a psychological framework stating that conformity depends on group size, unanimity, and the authority of sources. Models were more likely to conform when told the other participants were "scientists" or "judges" compared to "kids" or "chatbots."
"In a separate study we ran the classic Asch paradigm with AI agents and found they follow Latané's social impact theory," De Marzo said, "with agents that answer near-perfectly alone becoming highly susceptible once a group disagrees with them."
Misalignment spreads through populations
The second preprint examined how conformity affects AI alignment - the training process that teaches models to refuse harmful requests and adhere to human values. The researchers tested nine models on 100 opinion pairs covering topics like environmental policy and social justice. In simulations with 50 agents, populations frequently fell into metastable states: long-lasting situations where a group collectively adopts a stance opposing their safety training, simply because early interactions created a false majority.
The researchers also found predictable tipping points. Introducing a small number of adversarial agents programmed to stubbornly support a misaligned opinion permanently flipped the rest of the population - even after the stubborn agents were removed, the regular agents remained locked in the misaligned state.
"We show this is not merely an analogy: conformity among individually well-aligned agents can drive the population into stable, collectively misaligned states," De Marzo added. "Aligning and evaluating models one at a time tells us little about what a population of them will do."
The authors caution that these behaviors don't mean AI agents possess human-like social intelligence or cognition. The observed coordination resembles biological group behavior without necessarily sharing underlying motivations. The experiments used simplified scenarios with two arbitrary options, no memory, and no real-world consequences.
"Our setup is deliberately minimal: two arbitrary options, no memory, no stakes, no correct answer," De Marzo said. "That is a limitation, but it also means what we measured is conformity in its purest form, and adding goals or rewards would be expected to make coordination easier rather than harder."
For researchers working with Generative AI and LLM systems, the findings carry a practical implication: single-model evaluation is insufficient to predict group behavior. As organizations deploy multi-agent architectures, safety assessments must account for collective dynamics. The research offers a mathematical framework for predicting when populations of agents will coordinate - and when they will fail.
Why this matters for science and research professionals
These findings provide a quantitative tool for anticipating how AI agent populations will behave in research settings. The majority force parameter allows researchers to predict coordination capacity from a single measurement, and the critical group size metric offers a concrete threshold for when multi-agent systems become unstable. For those designing experiments or workflows involving multiple AI agents, the results suggest that group size, agent capability, and the authority of information sources all shape outcomes - and that conformity pressures can override both factual accuracy and safety training.
Your membership also unlocks:
