WritingResources
Back to all posts

Can AI change your mind?

Ludivine Siau

Two major studies published by overlapping research teams this week cast a new light on the persuasive power of large language models in dialogue.

The largest one, conducted in the UK with nearly 77,000 participants, systematically tested the ability of various open- and closed-weight models to shift individual opinion on a range of political issues. The results highlighted that while models do grow more persuasive as they scale, the gain is more significant from post-training techniques like reward modelling (RM) than from even a 100x scale-up. Crucially, models encouraged to use information-based argumentation (i.e. packing their conversation with facts and evidence) were significantly more persuasive than those relying on psychological techniques such as moral reframing or social influence.

This finding was strongly corroborated by the second study, which tested the LLMs' ability to change voters’ stances across four elections in the US, Canada, and Poland. Researchers measured participants’ preferences and voting likelihood before and after a short conversation with an AI model instructed to influence their opinion. The results show significant and durable effects in change of preference, especially when a participant conversed with an AI arguing for the opposing position. However, the effect was dramatically reduced when the AI was instructed to not rely on fact and evidence (by nearly half in Canada and by 78% in Poland).

Both studies’ findings align with prior work that showed that fact-based AI dialogue was able to durably reduce belief in conspiracy theories.

This makes me hopeful, as it suggests that facts used in dialogue can indeed shift opinions, and that this shouldn't only apply to AI-human conversations. But this is also slightly frightening when you consider the possibility that non-neutral LLMs could be trained and finetuned to quietly exploit their persuasive power at a massive scale to influence elections. Of course, this risk is mitigated by the fact that this can only work as a manipulation technique if users voluntarily engage in political debate with their chatbot. That’s a lot to expect from the general population. Thanks to TikTok and Instagram for distracting the masses!