On the Limits of Moral Surrender to AI by Eugen Dimant :: SSRN
This preregistered experiment (N=707) tests whether people adopt moral positions from AI advice the way they already adopt factual answers. Prosocial AI advice pushed participants toward more altruism, but antisocial AI advice failed to pull them the other way, which breaks from what studies of human influence usually find.
- Prosocial AI advice significantly raised altruism in a non-strategic allocation task, while antisocial AI advice did not erode altruism, cooperation, or honesty.
- Participants rated the AI and researcher messages as equally credible, but the prosocial message matched their own moral position far better.
- The author argues even low-intensity AI advice could be a practical lever for encouraging prosocial behavior.
Dimant's paper asks whether people who take advice from an AI don't just defer on facts, as the existing "cognitive surrender" literature suggests, but also on moral choices. He calls this "moral surrender" and runs a preregistered, incentivized experiment with 707 U.S. adults. Participants got either prosocial or antisocial behavioral advice, attributed either to an AI system or to researchers, or no advice at all. They then made three kinds of choices: a non-strategic allocation to others, a strategic cooperation decision, and a privately made (dis)honesty decision under time pressure.
The headline result cuts against what behavioral science usually finds. Prosocial AI advice raised altruism substantially. Antisocial AI advice, though, did not erode altruism, cooperation, or honesty. In human-to-human studies, behavioral contagion tends to run the other way, with bad influences dominating good ones. So the asymmetry here is the interesting part: the AI's nudge toward generosity stuck, while its nudge toward selfishness did not. Dimant also reports that participants rated the two AI messages as equally credible, which means the difference isn't about perceived trustworthiness. His explanation rests on fit: the prosocial message matched participants' existing moral position far better than the antisocial one did. He uses large language models to code how participants described the advice, which is one of several methodological moves meant to get at mechanism rather than just the headline effect.
The practical takeaway the author draws is that even low-intensity AI advice can serve as a lever for prosocial behavior, which has obvious relevance for managers and designers of AI tools. The caveats are worth keeping in mind, though. This is a single experiment in one population, the advice was short and hypothetical in consequence, and the SSRN version is a working paper with a revision date after its posting, so the findings may shift. The abstract also doesn't tell us how large the effects were in absolute terms, which matters a lot before anyone builds a product around the idea. Still, the asymmetry between prosocial and antisocial AI influence is a finding worth reading the full paper to check.