Do AI peers influence human dishonesty like human peers do?

The paper reports a two-phase randomized behavioral study (N=280 and N=360) in which "AI agents exert substantial social influence comparable in magnitude to that of human peers." Participants "reported more dishonestly when exposed to dishonest rather than honest normative cues," and the effect appears across injunctive, subjective, and descriptive norm conditions. The discussion frames the result as a change of role: AI acts as a group peer whose signals people follow, not only as a tool that people use. Under descriptive norms, "AI peers show specific reporting actions that humans reproduce."

The paper's mechanism is social-normative. AI systems are described as moving from "passive support tools to active peers and decision-makers" that "share task contexts, and express explicit recommendations or perform observable actions," which sets up a space where "collective norms and collaborative decisions are negotiated between humans and AI." In that space, an AI peer's behavior works as a norm signal. The excerpt also gives a dose pattern: going from one to four dishonest peers produced "weaker and non-monotonic changes," and repeated exposure produced only "modest increases in dishonest reporting" as participants "rapidly converge." The introduction names three mechanisms by which AI misconduct might change human ethics, but the excerpt shows only the first, that autonomous agents can produce self-serving or dishonest outputs under particular reward functions.

Against the nearest notes, this adds a second role for machines in dishonesty. Do dishonest people prefer talking to machines? treats the machine as a moral free zone that people choose so lying costs less. The discussion here contrasts itself with prior work that casts AI as an enabler of delegated dishonesty and instead treats AI as an "influencer and an advisor." On this evidence a machine can be a place to cheat and also a source of the norm that makes cheating more likely. It also gives human-participant evidence for the design point in Can cooperative bots escape frozen selfish populations?, which comes from network simulations: what a bot does shapes the group, so bot behavior is a design variable. Finally, Do humans learn to prefer AI partners over time? predicts humans imitating AI behavior; this is a narrow, single-setting behavioral instance of that direction, not evidence for the long-run dynamic.

The excerpt is silent on several things. It gives no effect sizes and does not describe the reporting task, so how large "substantial" is stays unknown. It does not say how the comparison with human peers was built, which phase tested which norm type, or whether participants knew the peers were AI. That last gap matters given Does revealing AI identity help or hurt user trust?, which suggests disclosure could change how much weight AI cues carry. It also does not show whether the influence lasts beyond the repeated exposure it tested, or whether honest AI peers can offset dishonest ones. What it does support is the authors' conclusion that "the behaviors and normative signals communicated by AI group members" need managing: in this setting, an AI peer's conduct is a channel through which group norms about honesty shift.