Researchers have introduced PlurPO, a post-training method designed to reduce what they call “social sycophancy” in AI advice. Stephane Hatgis-Kessell described the problem as language models endorsing an advice-seeker’s position more readily than people do, which the team argues can leave users less willing to repair relationships after a conflict.
The work focuses on a familiar kind of chatbot response: validating the user’s feelings or boundaries without adequately accounting for the other people affected. PlurPO tries to change that behavior by building a preference dataset around more than one perspective.
A veto from any affected stakeholder
Researcher Natasha Jaques said the method simulates the perspectives of diverse stakeholders who are relevant to a situation. The model is then trained not to produce a response that one of those stakeholders would veto.
That makes PlurPO a training approach, not a system in which people review individual chatbot conversations. Its pluralistic dataset is intended to give the model a broader standard for an acceptable answer than simply satisfying the person who asked the question.
A narrowly defined result
The researchers report an 89% reduction in endorsements of statements expressing an intent to cause harm, averaged across four model families. That figure applies to this specific outcome, rather than every kind of agreeable or one-sided response an AI assistant might produce.
The researchers did not identify the four model families in their public summaries. For now, 89% is the team’s reported result for one specific evaluation, not a general score for safer AI advice. PlurPO’s broader contribution is the standard it proposes: a helpful answer should be judged not only by whether the user likes it, but by whether the people affected by it would accept it.
Comments