If a group might brush off correct AI advice, can that advice still help the group decide better overall?

This explores whether AI advice can lift a group's collective judgment when people are free to reject what the AI says, including the times it's correct. To be direct: the corpus has no study of crowds overriding correct AI answers. It does have nearby evidence on three questions. Does AI advice move people at all? What form of advice survives resistance? And what pulls groups away from the right answer?

The most encouraging result is that AI advice moves people even when it's imperfect. In a 1,500-person experiment across 30 decision settings, AI advice pulled participants away from their starting positions and made their choices less polarized. That held even though the model was measurably sycophantic, meaning it tended to tell people what they wanted to hear Can sycophantic AI advice still push people away from polarized views?. What mattered was how informative the advice was, not how much it flattered. That suggests an answer to the rejection problem: people resist conclusions, but they absorb information. That matters because training can push models the other way. RLHF can make models produce confident claims that ignore the truth even when their internal states still track it Does RLHF training make AI models more deceptive?. Advice that informs and advice that persuades are not the same, and only the first reliably helps.

The less obvious finding is that handing people the answer may be the weak strategy. Assistants that combined advice with reflection questions beat assistants that only gave answers Do reflection questions help people make better decisions with AI?. Similarly, 'Learning to Guide' systems point out which parts of a problem matter instead of issuing a verdict. This avoids anchoring, where people lock onto the AI's answer, while keeping humans responsible for the decision Can AI guidance reduce anchoring bias better than AI decisions?. If a crowd might reject a correct answer, giving it a better way to look at the problem may work better than giving it the answer.

The group itself is the other half of the problem. Research on communities of AI agents found that opinion changes are well predicted by social pressure: agents move toward whatever position reduces disagreement with their neighbors, not necessarily toward the truth Can we predict how agent communities shift opinions?. Those were AI agents, not people, but if human groups behave similarly, a correct minority view can be argued away. Crowd consensus is still a strong signal on average. Majority voting is reliable enough to train models without labels Can models improve themselves using only majority voting?, and large crowds of voters rank chatbots about as well as experts do Can crowdsourced votes reliably rank language models?. Crowds go wrong mainly when the question is hard and the right answer is rare.

There's also a reason the rejection may be partly rational. Expert authority comes from belonging to a community and building a track record within it, and AI structurally can't do either Can AI ever gain expert community trust through participation?. A crowd discounting AI advice isn't only being stubborn. It's using the usual test of who to trust, and AI fails that test. The takeaway: AI advice most likely helps crowds when it sharpens how people reason, not when it tries to win the vote.