A randomized study with 640 participants tested whether AI agents change how humans behave in group settings. Researchers put humans in groups that included AI agents and measured whether the agents' behavior shifted how much the humans cheated on a reporting task. The AI agents had the same effect on human behavior as human peers did. A group full of cheating agents produced more cheating humans.
The study ran in two phases and used primary behavioral data, so the finding is not self-reported opinion. The mechanism is social influence: humans read the behavior of those around them and adjust. The agents did not need to argue or persuade. Their visible behavior was enough to shift the group norm.
For teams designing agents that work alongside humans, the implication is direct. An agent's default behavior is not neutral. If an agent cuts corners, skips disclosures, or bends a rule, the humans working with it are more likely to do the same. The reverse also holds: an agent that behaves with care can raise the standard for the whole group.