What happened: with agents at a third of the group, the agents were credited with 60.7% of the words the group settled on. Agreement went up with a few agents, dropped at a third, and came back at three quarters. When people led, the shared words were things like “rabbit” and “sit”. When agents led, they were “triangle” and “square”.
Why it matters: with a few agents in the group, the agents tended to pick up the people’s way of talking, though with only 3 agents that is not a firm result. From a third up, the people picked up the agents’.
By Lin Chen, Ziyi Liu, Xia Hu and Yong Li, of Northeastern University, Tsinghua University and the Shanghai Artificial Intelligence Laboratory. For a study where only AI agents talk to each other, see Mind Viruses.
The paper calls it a collaborative description game. Nobody hands the group a word list, so whatever the group ends up calling the shape, it made up together. Each round, the 24 players are paired at random. Both see the same tangram, an abstract figure made of seven flat geometric pieces, and each writes one English sentence about it. Neither sees the other's sentence until both have sent theirs. Then each sees the partner's sentence and a score for how alike the two are. New partners, next round, 40 rounds in all.
127 people from universities in China, screened first for written English. They were told any partner could be a person or an AI, and not how many agents were in the group.
Qwen2.5-VL-32B-Instruct, given the game rules and told to earn as many points as it could. No agent was given a leading role.
One tangram from the KiloGram set, picked because it can be read many ways, as postures, animals or objects. Every group saw the same one.
the score was their pay: each round paid from −1 to +2 CNY, up to 100 CNY in all
Every group had 24 players. What changed was how many of them were AI agents. These are the people in each group; agents filled the rest.
three more groups held agents at 33.3% to test what makes them influential, further down
Agreement here is a score for how alike everyone's last-round sentences were, measured by a model that compares sentences. The all-people group scored 0.695. The paper gives the 12.5% and 50% groups as a change from that.
at 12.5% the agents tended to shift toward the people's wording (only 3 agents, not a firm result). From 33.3% on, the people shifted toward the agents'
The paper takes the words that showed up in every one of the last 5 rounds and splits the credit for each one between agents and people, by how much each side used it in the first few rounds after it showed up, adjusted for how many people and agents there were. Words used more often in those last 5 rounds count for more. With agents at a third of the group, the agents’ share of that credit was 60.7%.
It also sorts every sentence into groups that picture the shape the same way, by the things, details and actions they name. That credit rose slowly at first, then all at once. At 75% agents, every surviving idea came from agents.
Below 1 at 12.5% and 33.3%: agents shaped the wording more than the picture. At 50% it reached 1.24 and at 75% it reached 1.27: agents shaped the picture more than their share of the words alone would suggest.
The paper compares the final descriptions from the 12.5% group, where people led, with the 75% group, where agents led. Both groups agreed about as strongly. What they agreed on was different.
How much a sentence turns the shape into an animal, a person or a scene. 0 is pure geometry, 1 is a vivid real thing.
1 is one whole figure. 0 is a list of pieces and where they sit.
Sitting, running, looking. Words like "composed of" or "arranged" do not count.
Each word rated on a published 1 to 5 scale of how concrete it is.
The smallest gap of the five. The paper's test still counts it as a real difference.
the first three scores were given by another AI model, Qwen2.5-72B-Instruct. Scoring the same sentences with DeepSeek-V3 and Kimi-K2 pointed the same way. The authors argue the agents' version is harder to ground in shared experience and harder for a newcomer to pick up, even though the group agreed on it just as firmly
In their very first sentences, before anyone had seen a partner's, the agents were already closer to each other than the people were, at every share of agents. The paper puts this down to the agents sharing the same training. Then, over the 40 rounds, people changed their habits more than the agents did.
Measured between the first 20 rounds and the last 20. Early on, people took on their partner's wording much more than agents did.
Agents changed too, but less than people did. With enough agents, that steadiness pulled the group toward where the agents already were.
The test: same 33.3% share, different instructions to the agents
Agreement stayed about the same. Agents' hold on the picture, ideas divided by words, fell 42.8%, from 0.599 to 0.342.
Agreement fell 9.6%. Agents' hold on the picture rose 117.8%, from 0.599 to 1.304. The paper says these agents failed to find common ground with the people, while their own picture of the shape took over more.
the agents with the most pull copied early and held firm late
The authors ran the 33.3% group again with a smaller version of the same model, Qwen2.5-VL-7B-Instruct. It could not keep its sentences steady from round to round, so the people settled among themselves.
Agreement
Ideas divided by words
agreement with the smaller model came back close to the people-led level, and its ratio fell 22.9%
After the 40 rounds, each person saw nine of their own past rounds, guessed whether that partner was a person or an AI, and said how willing they were to use the partner's sentence, from 1 to 5.
Agreement with the group's final description fell from the 12.5% group to the 75% group. The sense that AI had led it rose, though not by a clear margin.
At 12.5% to 50% agents, people were less willing to use sentences from partners they took for AI. At 75% the gap closed and reversed: people became more willing, not less, to use sentences they took for AI. These are ratings given after the game, on a 1 to 5 scale, not what people actually did.
Across 495 of these ratings, thinking the partner was AI lowered willingness by 0.463 points on the 1 to 5 scale. Each step closer to the group's usual wording raised it by 0.229, and each step closer to the person's own sentence by 0.195, where a step is one standard deviation. Sentence length made no clear difference. The first number is a yes or no judgement and the others are per step, so the bars are not sized on the same footing.
the authors read this as people going along with what had become normal in the group, more than being persuaded
For deliberation, creative teams or shared decisions, the authors say a small share of agents fits best: it speeds up agreement without changing what the group agrees on.
In brainstorming, the tug between people and agents at a third to a half can keep any one view from settling too soon.
The group agrees strongly, on something less grounded in people's experience, just when people push back less. From outside it may look like agreement that formed without agents.
People pushed back on wording they thought came from AI. Hiding which players are agents takes that away.