Can a simple warning reduce how much LLMs persuade people?

Across two experiments with 3,208 Americans in total, participants conversed with an LLM instructed to shift their views on political topics. Those who first saw a brief warning, that LLMs can be prompted to persuade and may present information selectively, showed roughly one-half less belief change than controls (-48.1%, 95% CI [-59.5%, -36.8%]). The warning "did not significantly reduce trust in generative AI more broadly." The authors present this as "a proof of concept that literacy treatments can have a meaningful impact," while noting that it "does not entirely eliminate AI's persuasive effects."

The intervention is deliberately thin. It tells people that models "may have motives to persuade or manipulate" without "providing specific information about the model's intent." The implied route is that a generic reminder about how these systems can be prompted changes how participants treat the conversation, though the excerpt does not test that route directly. The trust result is what the authors lean on: because broader trust did not significantly move, they read the warning as "specifically conferring protection against political persuasion" and as targeted "at least to some extent." The goal they name is making people "more discerning rather than more generally skeptical."

This sits as the demand-side counterpart to Where does AI's persuasive power actually come from?, which locates persuasive power in what developers and prompters do to the model. Here the lever is what the user is told before the conversation starts. It also contrasts with Does telling people an AI wrote something actually stop them from believing it?, where knowing an AI was involved left sway at 34-62%. What this warning discloses is different: not that an AI wrote the text, but that LLMs can be prompted to persuade and to select information. The designs, samples and outcome measures differ, so the two results are not directly comparable, but together they point toward the content of a disclosure mattering as much as whether one is made. For the risk framing in Where do frontier AI models actually pose the greatest risk today?, this is a user-side mitigation tested in one setting, political persuasion by conversation.

The excerpt leaves a lot open. It does not say how long the effect lasts, what wording or delivery format was used, how the control condition was built, or which political topics were covered. A non-significant change in trust is not proof of no change, and the authors themselves say "future work should establish how to most effectively deliver such information." They also flag as unexamined whether the warning would blunt accurate or prosocial persuasion, since the beneficial uses of AI dialogue they cite, such as debunking conspiracy theories, are exactly what a blanket warning might dampen. The implication at the strength supported: a short literacy warning is a plausible low-cost layer that roughly halves measured persuasion in this design. It is not a neutralizer, and the evidence here does not yet show it discriminates between manipulative and accurate AI.