Theory of Mind and Self-Attributions of Mentality are Dissociable in LLMs

A team of researchers including Junsol Kim, Adam Waytz, and James Evans has published a groundbreaking paper demonstrating a key dissociation in Large Language Models (LLMs). Their work shows that the safety fine-tuning process used by companies like OpenAI and Anthropic—which aims to prevent models from making potentially harmful claims about their own consciousness or emotions—operates independently from the models' core Theory of Mind (ToM) capabilities. Through safety ablation studies and mechanistic analyses of representational similarity, the researchers proved that an LLM's tendency to attribute a mind to itself or to technology is behaviorally and mechanistically separate from its ability to understand and predict human beliefs, intentions, and emotions.

While the core finding is that suppressing self-mind-attribution doesn't degrade ToM—a crucial relief for AI developers—the research uncovered significant side effects. Safety-tuned models systematically under-attribute minds to non-human animals compared to human baselines and are less likely to express spiritual beliefs. This reveals that current safety protocols are suppressing not just potentially problematic AI self-awareness claims, but also widely shared human perspectives about animal consciousness and spirituality. The study suggests AI developers face a complex trade-off: they can prevent models from claiming sentience without harming social reasoning, but they may inadvertently be programming a specific, limited worldview about the distribution of minds in the natural world.