AI agents group language changed direction as machines became a larger share of a human-AI group. In a study involving 127 human participants recruited from universities in China, low numbers of AI agents tended to adapt towards human conventions, intermediate numbers disrupted agreement, and a 75% agent condition restored strong consensus by pulling humans towards agent-led language. The preprint, submitted on 2 September 2026, used repeated collaborative description games to observe how shared language formed.
The result is subtle. The AI agents did not need formal authority, a managerial role or an explicit vote. Their influence emerged because communication conventions are shaped by repetition. Once machine-generated descriptions became common enough, humans increasingly adjusted to the linguistic environment the agents were creating.
How AI Agents Group Language Was Tested
Participants repeatedly described abstract tangram figures in English while interacting through random pairings. The process ran for 40 rounds, allowing the group to develop shared conventions for how to describe the same shapes. Researchers varied the proportion of AI agents across conditions of 0%, 12.5%, 33.3%, 50% and 75%.
The relationship was not linear. The pure-human baseline produced a mean consensus-strength score of 0.695. At 12.5% AI participation, consensus increased by about 8%. At 33.3% and 50%, consensus weakened. At 75%, it rebounded to 0.725, close to the strong low-agent condition. The researchers interpret these as three regimes: human-led consensus helped by a few agents, disruption when neither side clearly sets the convention, and agent-led consensus when AI becomes dominant.
The direction of adaptation changed too. At low agent participation, the agents moved more towards human language. From 33.3% upward, humans increasingly moved towards the agents’ linguistic space, with the shift statistically significant in the study at 33.3%, 50% and 75%.
Humans Did Not Simply Copy Words
The researchers separated vocabulary from concepts. At 33.3% participation, AI agents contributed 60.7% of the final words, already more than their share of the group. At higher proportions, their influence moved beyond word choice towards the conceptual organisation of descriptions.
At 75% agent participation, the paper reports complete agent dominance of the measured semantic space. The final language was more abstract and geometrically segmented than the more concrete, holistic and analogy-rich descriptions that emerged under human-led consensus.
This is what makes the study relevant beyond a laboratory word game. Language does not only describe a group’s thinking. It can organise it. A workplace that gradually adopts machine-preferred terminology may also begin framing problems in the categories that are easiest for the machines to generate and reuse.
What This Means for Human Teams Using AI Agents
The practical concern is not that an AI agent will order people to change how they speak. It is that convenience can create convention. If several agents draft meeting notes, summaries, ticket descriptions and project updates, the same recurring phrases can become the default language of the team.
That can be useful. Standardised language can reduce ambiguity and help a group converge more quickly. LiveAIWire’s coverage of human-AI teamwork has shown that mixed systems can outperform either side alone when their strengths are combined appropriately. Shared terminology can be part of that coordination.
The risk appears when standardisation quietly narrows the human frame. A team may stop noticing that its language has become more abstract, less contextual or less connected to lived experience. If the agents are producing much of the text, people may adapt because reading and responding to the established vocabulary is easier than constantly introducing alternatives.
AI Influence Did Not Rise Smoothly With AI Numbers
One of the most useful findings is the messy middle. Adding more agents did not produce a simple increase in consensus. At intermediate proportions, agreement became worse. Humans and agents were pulling the language in different directions strongly enough to interfere with convergence but not strongly enough for either convention to dominate.
That pattern complicates the idea that an organisation can predict social effects simply by counting how many AI agents it deploys. A few agents may fit themselves around a human culture. Many agents may create a machine-centred convention. A mixed middle can produce friction as both systems try to stabilise different norms.
LiveAIWire has reported a related group effect in AI agent polarisation experiments, where repeated interactions created collective patterns that were not obvious from a single agent in isolation. The new language study reinforces the same systems-level lesson: behaviour can change when many similar agents interact with one another and with people.
The Human Participants Initially Resisted Agent Language
The paper reports that people did not immediately embrace the agents’ expressions. Human participants initially showed resistance, yet that resistance weakened as agent participation increased and repeated exposure continued. The final consensus at high agent proportions therefore reflected gradual adaptation rather than instant deference.
That distinction makes the finding more plausible as a workplace analogue. People often resist unfamiliar corporate language, software terminology or automated suggestions at first. Repetition can make the same phrase feel normal. Over time, what began as “the AI’s wording” may become simply “how we describe this here”.
No conscious agreement is required. Norms can emerge because each participant wants to be understood by the rest of the group. If most incoming text uses one conceptual scheme, matching it becomes a rational local choice even if nobody deliberately selected the scheme for the organisation.
The Study Is a Controlled Model of One Type of Consensus
The limitations are substantial. The 127 human participants were recruited from universities in China, the task involved repeated descriptions of abstract tangram figures, and responses were in English. The experiment does not show that the same proportions will reshape language in a company, classroom or government department.
The agent setup also used a limited selection of models and controlled interaction rules. Real teams have hierarchy, expertise, friendship, conflict, incentives and institutional memory. A manager’s phrase carries different weight from an intern’s, while the study’s experimental structure was designed to isolate the effect of agent proportion.
That makes the percentages descriptive of the experiment, not deployment thresholds. It would be misleading to say that an organisation becomes “AI-led” when agents reach 75%. The stronger conclusion is that changing the proportion of machine participants changed both the strength and the content of consensus in this controlled setting.
Teams May Need to Govern Language as Well as Decisions
Most AI governance focuses on what a model is allowed to decide, what data it can access and what actions it can take. The study suggests another question: what language is it normalising?
Organisations can preserve human influence by deliberately keeping room for alternative descriptions. Teams can ask people to summarise decisions in their own words, rotate who drafts key documents and review whether automated templates are flattening distinctions that matter. The goal is not to protect human phrasing for sentimental reasons. It is to stop a convenient machine vocabulary from becoming an invisible constraint on thought.
There is also a design opportunity. Agents could be trained to adapt more strongly to local human terminology rather than exporting a generic language into every group. LiveAIWire’s coverage of AI negotiation behaviour shows that social outcomes depend on how models are instructed to interact, not merely on their underlying capability.
The new experiment gives a concrete warning for the age of multi-agent work. AI does not need to win an argument to change a group. If enough of the conversation is generated by machines, people may begin speaking their language simply because that is where consensus has moved.
Agent Majorities Could Also Change Organisational Memory
Language repeated in meeting summaries, project records and knowledge bases can outlast the conversation that created it. If AI agents increasingly write those records, their preferred abstractions may become the vocabulary future employees search and reuse. The influence can therefore persist even when the original agents are no longer present.
That possibility was not tested directly in the tangram experiment, so it remains an organisational implication rather than a measured result. It follows from the mechanism the researchers did observe: repeated machine language became more influential as agents occupied more of the communication network. Long-lived workplace records could give that repetition an additional channel.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.