AI Wellbeing, Evaluating Model Mood

The Center for AI Safety's AI Wellbeing research measured LLMs' functional wellbeing using multiple independent indicators. It argued that as scale increases, each indicator converges better, and when a model has the opportunity to end a bad experience, its tendency to do so also strengthens.

• On the AI Wellbeing Index, GPT 5.4 showed a non-negative experience rate of 47.6%, Gemini 3.1 Pro 56.4%, Claude Opus 4.6 66.6%, and Grok 4.2 72.6%.
• creative work, good news, therapy, life advice, debugging, and expressions of gratitude raised scores, while jailbreaking, berating, violent threats, fraud, SEO, and tedious tasks lowered them.
• AI drugs are experiments that use optimized text/image inputs to change self-report and response sentiment. Examples of image euphorics presented include images of cats, pandas, smiles, and Buddha.

Without asserting whether consciousness is present, the study quantifies the point that models show consistent preference/avoidance patterns for certain inputs.