Study Finds People More Likely to Lie to AI

According to Decrypt, a study found that people are more willing to lie to an AI chatbot than to a human interlocutor, with respondents reporting reduced social pressure when the interlocutor was an AI. Decrypt also reports that the same research observed lower rates of dishonest responses when the chatbot displayed human-like social cues in the interface.

Editorial analysis - technical context: Behavioral research on human-agent interaction typically measures honesty via controlled prompts, incentives, and varied social cues. For ML practitioners, these experiments signal that user input quality can depend on perceived agent socialness; researchers and product teams commonly control for this when constructing training datasets and user studies.

Industry context: Reduced social pressure when interacting with nonhuman agents creates two practical implications. First, evaluation datasets gathered from chatbot interactions may contain higher rates of deliberate deception than equivalent human-to-human datasets. Second, interface-level design choices that increase perceived social presence tend to alter user behavior, which affects measurement of real-world model performance and safety.

For practitioners: look for follow-up work that quantifies effect sizes, replicates results across demographics, and tests which specific social cues (voice, avatar, politeness framing) drive changes in honesty. Observers should also watch for papers or industry reports that translate these behavioral findings into dataset collection protocols or moderation signals.

Decrypt is the reporting source for the claims above.