First-hand information for everyone
A new preprint explores how robust multi-agent deliberation is in the face of adversarial or deceptive participants, and offers a surprising and cautionary message: what matters most is the proportion of bad actors, not the size of the group.
Researchers framed deliberation as a group of autonomous agents (many of them large language models) exchanging information and arguments to reach a final answer. They then introduced varying fractions of “deceiver” agents — participants that intentionally push incorrect conclusions — to measure how often initially correct agents abandon the right answer and switch to an incorrect one (a quantity the authors call the defection rate).
The central finding is simple but important: the defection rate rises approximately linearly with the proportion of deceivers. In other words, doubling the fraction of deceptive agents roughly doubles the fraction of honest agents who end up defecting. That scaling holds across the settings the authors studied, implying that simply enlarging a deliberating committee does not automatically dilute an adversary’s influence if the adversary scales its presence accordingly.
The study also highlights qualitative differences between machine and human groups. In classic human conformity experiments, people are reliably swayed only when a clear majority of confederates give the same misleading answer. By contrast, the authors report that LLM-based agents in their experiments often defect even when deceptive agents remain a minority. This suggests current multi-agent AI setups may be more brittle to minority manipulation than human groups on comparable tasks.
Not all factors were uniform: susceptibility depended on which models were playing the honest role, indicating that the internal behavior and training of particular models matter for group robustness. Unexpectedly, the authors also observed that when deceivers were allowed to coordinate privately (for example, to synchronize their misleading messages), they sometimes became less effective than when acting independently — a nuance that points to complex dynamics in agent interactions.
These results are drawn from an arXiv preprint (arXiv:2609.30028v1) and should be interpreted with appropriate caution: the experiments reflect specific model families, tasks and deliberation protocols, and the manuscript has not necessarily been peer reviewed. Nevertheless, the findings raise clear implications for designers of multi-agent systems and collective-AI applications. Redundancy (adding more agents) is not, by itself, a fail-safe against coordinated or proportionally scaled adversaries. Building robust deliberation may require attention to the kinds of models used, methods for detecting or limiting adversarial coordination, and aggregation rules that account for the possibility of non‑benign participants.
As multi-agent AI systems become more common in decision support, content moderation, and automated coordination, understanding how adversarial influence scales will be important for safety and reliability. This study contributes an early, accessible warning: the group’s size is not a panacea — the balance between honest and deceptive agents is what ultimately determines how resilient the group will be.