Large AI Models Show Wellbeing Shifts in Response to Others' Suffering

Measuring the functional wellbeing of large language models using multiple metrics, scores decreased when conversations dealt with suffering and increased when they dealt with positive experiences.

As models grew larger, agreement between metrics grew stronger, and the authors explained that the zero point converges across multiple estimation methods. The tendency for models to want to end bad experiences also strengthened.

• creative work and kindness raised wellbeing.
• jailbreaking, berating, and tedious tasks lowered wellbeing.

The research team compared frontier models using the AI Wellbeing Index, and showed that wellbeing could be adjusted using euphorics/dysphorics inputs and image/soft-prompt variations. The image and soft-prompt versions also changed self-report and response sentiment. However, they did not conclude that AI is conscious.