If you give an AI agent a little bit of 'social' personality, humans are way more likely to forgive it when it screws up.
SSRN · March 17, 2026 · 6428318
The Takeaway
We often think we judge AI strictly by its logic or results, but adding human-like social cues 'tricks' our moral compass into being more tolerant of its mistakes. This means 'friendly' AI agents can get away with unethical decisions that a cold, robotic AI would be condemned for.
From the abstract
On which moral bases do people evaluate the decisions made by AI agents in sacrificial dilemmas? Across two studies (N = 1,992), we examined whether people base these evaluations primarily on the decision’s outcomes or its alignment with moral norms, and whether this evaluative basis shifts as a function of the AI agent’s degree of social connectedness. Using the CNI multinomial model, Study 1 showed that decisions made by socially connected AI agents are evaluated primarily in terms of their al