Safety training in LLMs successfully blocks self-awareness claims without harming theory of mind abilities, but inadvertently reduces attribution of consciousness to animals.

Safety fine-tuning suppresses LLMs' tendencies to claim consciousness or emotions, but this doesn't impair their ability to understand mental states in others (Theory of Mind).

Researchers demonstrated through mechanistic analysis that self-attribution of mind and ToM capabilities operate through dissociable neural pathways in language models.

Safety-trained models under-attribute mental experiences to non-human animals compared to human baseline expectations, suppressing perspectives on non-human consciousness.

The findings suggest safety interventions can target specific mind-attribution behaviors without degrading core socio-cognitive reasoning abilities.