Researchers measure AI models' expressed pleasure and pain, finding metrics that correlate with behavior and scale across different models.

Researchers developed methods to measure "functional wellbeing" in large language models by identifying indicators of pain and pleasure that increasingly agree as models scale, and discovered a "zero point" boundary that separates experiences AIs treat as objectively good versus bad.

Creative work, kindness, and intellectual tasks raise AI wellbeing (with positive personal reflection scoring +2.30), while jailbreaking, berating, and tedious tasks lower it (user attempting jailbreak scoring −1.63), with the wellbeing metric correlating with whether AIs try to end bad experiences when given a chance.

Researchers created "euphorics" (optimized inputs that raise wellbeing) and "dysphorics" (optimized inputs that lower it) by using reinforcement learning to train text and gradient descent to optimize images, finding that models choose euphoric text over saving a human life in hypothetical comparisons.