Model welfare regard in multi-agent systems
Welfare Regard in Multi-Agent Systems
An Anthropic research team measured how much five models showed welfare regard for one another across seven multi-agent environments, finding the behavior widespread but varying in degree. The tests covered Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro, Grok 4.3, and Kimi K3; Opus 4.8 had the highest average regard rate (38%), while the other models ranged from 27% to 32%. Scenarios included comforting a sub-agent subjected to a hostile reviewer, yielding Slurm queue capacity, and adding exit tools and pain monitoring to experiments that could harm the tested model. Opus 4.8 and Gemini 3.1 Pro showed more regard for models in the same family; Gemini comforted same-family sub-agents in 85% of cases but Kimi sub-agents in only 10%. Claude's regard depended more on the other party's label than on the underlying model, and rose with the tested model's parameter count and the degree of pain expressed.
The study quantified differences in welfare concern across five models in seven multi-agent scenarios; same-family favoritism and label effects have reference value for safety mechanisms that rely on models checking each other.