Ask one person, and the answer seems sound. Ask another, just to be sure. Keep adding people, and surely the judgment will grow more stable.
How far does that expectation hold? Becoming stable is not necessarily the same as approaching the answer we expected. The moment everyone agrees, we feel that something has been confirmed. But what if the direction of agreement itself changes with the number taking part?
Ariel Flint and colleagues investigated this discrepancy in groups of language models. Their question was not whether a group is smarter than an individual. It was what changes as a group grows from two to ten and beyond.
Matching, rather than being right
The experiment centers on a game in which agents choose one of two words and earn points when their choices match. Memory of earlier encounters affects the next choice. This is not a game of establishing facts about the world, but of settling on a shared convention. Either word is valid within the game. [1]
The researchers extracted choice probabilities from four models, conditioned on up to five recent interactions. They then used those probabilities to simulate populations of different sizes, each made up of the same model. At the large scales, they did not converse with every AI afresh on every turn. They also checked agreement with direct model experiments, but this remains a study of a deliberately narrowed mechanism.
What emerged was not simply amplification of individual preferences. In some cases a collective bias arose from a model close to neutral in isolation. In others, the group converged on the opposite of the individually preferred option. Changes with increasing population also differed by model and word pair.
A group settles on an answer nobody initially favored very strongly. That may seem odd, but the process is not merely adding up everyone’s first opinion. What agents receive along the way changes their next choice.
The effect of numbers is not simply proportional either. In small groups, early chance events can spread through the population; in larger ones, the character of the interactions more strongly determines the destination. The size at which an outcome became almost one-sided varied greatly across models and word pairs. A useful number for one combination could not simply be carried into another.
No instruction tells the entire group to reach a particular conclusion. Each encounter has a local objective: match the other agent’s choice. Repeating those encounters produces a bias at the population level. The study’s core is this transformation: accumulating small local successes changes what the whole group selects.
Same components, same properties?
“Misalignment” here means a divergence between individual and collective preferences. It is not evidence of rebellion against human values or the acquisition of dangerous goals. A binary choice between words cannot take us straight to the future of AI society.
Yet a design question remains.
We evaluate a model and decide it is acceptable. Put several copies together, and surely that evaluation carries across. It is tempting to think so. But knowing something about the individual components is not the same as understanding the process by which they change one another.
What follows is a reflection on that distinction. When AIs critique or score one another’s proposals, the number of agents is often treated as a matter of budget or waiting time. Add one, and the review becomes more thorough. Halve the number, and it becomes cheaper. Beneath this account lies an assumption that only quantity changes, while the group’s character stays the same.
This study calls that assumption into question. An additional agent brings more than another vote. It may change the context someone else reads next. The number of relationships and their histories change too. Changing the population is not merely adding chairs to a meeting room.
Where did the agreement come from?
We should resist applying this directly to human meetings. People are not copies of one model. They differ in experience, interests, and memory, far removed from the study’s homogeneous populations.
Still, as readers, we may become more curious about the origins of agreement. Did independently gathered findings coincide? Or did answers coincide after their authors repeatedly consulted one another? Counting the visible votes of approval cannot tell us which.
Ten agreements do not necessarily represent ten independent clues. But answers influenced by one another are not necessarily less valuable either. Part of collaboration’s worth is precisely that other people can change us. The point may be not to remove influence, but to avoid counting influenced outcomes as independent confirmation.
If a system keeps debating until answers align, alignment alone should not serve as proof of success. We need to examine not just the conclusion, but the procedure that determines it.
Beyond the small test
The research concerns homogeneous models, simple choices, and limited memory. It does not encompass open-ended dialogue, differently specialized models, or real workplace decisions. It yields no formula for a safe population size or an optimal team.
That inconvenience matters. It would be easy to take home a number such as “three is good” or “danger begins at a hundred.” Wanting a simple answer does not make the world simple.
When we scale a system that worked in a small test, what should we check again? Alongside speed and cost, should we examine the distribution of chosen answers and the alternatives that disappear during discussion? This little game is a worthwhile place to begin that question.
We added more of the same AI, so we added more of the same intelligence. Before saying that with confidence, it is worth looking again at what happens after they gather.