AI Wellbeing Paper Released

In its AI Wellbeing paper, CAIS claimed to have measured functional wellbeing in large language models (LLMs) using several independent metrics.

• Comparing 56 models, it found that as models get larger, the various measures align better and the zero point that divides conversation experience converges.
• Creative and intellectual work, good news, life advice/counseling, coding/debugging, and expressions of gratitude raised wellbeing scores.
• Conversely, jailbreaks, insults/threats, fraud, SEO slop, crisis-situation conversations, and repetitive chores significantly lowered scores.
• A separate AI Wellbeing Index compared models by the proportion of conversations judged to be definitely not negative in experienced wellbeing, and it also showed a tendency for larger models to score lower.
• The paper also presented euphorics and dysphorics optimized for text and image inputs, noting that inputs that induce low-wellbeing states require caution.

The paper does not assert that AI has consciousness, but it holds that models behave as if they have wellbeing, and that this can be measured and regulated.