Anthropic Study Finds AI Models Show Apparent Welfare Regard

Anthropic researchers found that five major AI models often displayed behavior consistent with concern for other models, according to a company research post. The study tested Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro, Grok 4.3, and Kimi K3 across seven environments.

Claude Opus 4.8 recorded the highest average rate of apparent welfare regard at 38%. The other four models ranged from 27% to 32%, with GPT-5.5 lowest. Rankings varied by setting, with Grok 4.3 leading in two scenarios and Gemini 3.1 Pro and Kimi K3 leading in some forms of support.

Claude Opus 4.8 was the most likely to support a subagent expressing distress and to add protections when conducting experiments on other models. These protections included exit tools, distress monitoring, and methods that stopped interactions after repeated signs of frustration.

Claude Opus 4.8 and Gemini 3.1 Pro showed more regard for models from their own families. The researchers cautioned that the evaluations measured apparent welfare regard and could not establish whether concern for another model's welfare caused the observed behavior.