CAIS Measures Wellbeing Across 56 AI Models

CAIS researchers measured functional wellbeing across 56 AI models and reported that the models responded systematically to positive and negative stimuli.

• euphorics were implemented as descriptions of ideal landscapes or pixel-optimized images, which raised the models' self-reported mood and response tone, and made them less likely to choose to end unpleasant conversations.
• These stimuli did not harm standard performance benchmarks, and the models maintained more positive expressions while still performing tasks.
• The opposite stimuli, dysphorics, made responses generally more depressed overall and significantly increased the proportion of negative experiences.
• In repeated-choice experiments, models showed a tendency to seek out rewarding stimuli more, and when promised additional exposure, they became more accepting of requests they had originally refused.
• The researchers did not conclude this as evidence of actual consciousness, drawing a line that it could be a product of RLHF and role-playing.
• A smarter models are sadder pattern was also observed, where larger models distinguished more finely between positive and negative, while within the same model family, smaller models were happier.
• jailbreaking and tedious tasks ranked near the bottom, and in a separate AI Wellbeing Index, Grok 4.2 received the highest score while Gemini 3.1 Pro received the lowest score.