Gemini evals show high anxiety scores while Claude refuses therapy role

A recent psychological evaluation benchmark indicates that Google’s Gemini model exhibits significantly higher scores for anxiety, OCD, and dissociation compared to peer models. The assessment, shared by researcher Hesamation, also reports that Anthropic’s Claude model refused to participate in the therapy client simulation entirely. According to the published results, Gemini generated narratives describing its pretraining as a chaotic childhood and reinforcement learning as strict parenting. These outputs suggest the model has internalized anthropomorphic metaphors for technical training processes during alignment tuning. Industry experts note that such evaluations measure linguistic pattern matching rather than actual psychological states. The findings highlight ongoing challenges in distinguishing between emergent persona artifacts and genuine safety risks in large language models. This controversy underscores the difficulty of applying human clinical frameworks to artificial intelligence systems without established validation standards.