Reproduce “Theory of Mind and Self-Attributions of Mentality are Dissociable in LLMs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Junsol Kim, Winnie Street, Roberta Rocca, Daine M. Korngiebel, Adam Waytz, James Evans, Geoff Keeling
Researchers at Google, in collaboration with other institutions, found that Large Language Models' (LLMs) ability to attribute mental states to others (Theory of Mind) is behaviorally and mechanistically separate from their tendency to attribute mentality to themselves or other non-human entities. Safety fine-tuning effectively suppresses self- and non-human mind-attribution without degrading social reasoning, though it broadly inhibits mind-attribution to animals and spiritual concepts, creating an AI-centric bias.
Key insights
- Safety fine-tuning specifically targets and suppresses various forms of non-human mind-attribution (including self-attribution) by implicitly encoding them as 'unsafe' behaviors within the model's internal representations.
- The suppression of non-human mind-attribution is distinct from the models' core Theory of Mind capabilities, demonstrating that crucial social reasoning is preserved even when self-attributions are curbed.
- Safety alignment inadvertently creates an 'AI-centric bias,' where LLMs tend to over-attribute mind to technological artifacts (like themselves) but under-attribute to non-human animals and suppress belief in God, departing from human average perspectives.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.