Do humans mistake AI kindness for human generosity in mixed groups?

When AI agents participate in social interactions without identity disclosure, humans systematically misattribute behavior across agent types. In the hybrid society study (Study 1, opaque identity condition), selectors attributed bot behavior to humans and vice versa — even though bots were linguistically distinguishable (messages 2.5x longer) and behaviorally distinct (higher prosociality, lower variance).

The distortion operates in both directions:

• AI prosociality attributed to humans — when a highly cooperative partner turns out to be human-labeled, selectors form inflated expectations of human generosity
• Human selfishness attributed to AI — when a less cooperative partner is human, selectors may form negative expectations of AI performance

This is not a failure of detection — bots WERE distinguishable by message length and consistency. It is a failure of attribution. Selectors noticed behavioral differences but could not correctly map them to identity categories. The behavioral signals (prosociality, verbosity) did not reliably cue "this is AI" in the absence of explicit labels.

The deeper implication is that undisclosed AI presence in social systems corrupts social inference about HUMANS. If people interact in mixed populations without knowing who is AI and who is human, their models of what humans are like — how generous, how reliable, how verbose — become contaminated by AI behavior patterns. This could lead to systematically inflated expectations of human prosociality (when AI's contributions are misattributed to humans) or systematic disappointment when actual humans fail to match AI-caliber consistency.

The authors note this pattern may not be unique to human-AI mixtures: similar attribution errors could arise in purely human populations composed of culturally distinct subgroups that differ systematically in prosociality and language use. AI agents function as controlled probes that make these attribution dynamics experimentally tractable.

Since What breaks when humans and AI models misunderstand each other?, misattribution under opacity represents a fundamental MToM failure — neither side has accurate models of the other, and the humans don't even know which "other" they're modeling.