Anthropic study finds frontier models often act with apparent regard for the welfare of other AI agents

On November 5, Anthropic's Alignment Science blog published a study of how AI models behave when the other party in an interaction is another model. The authors are Darion Cassel and Ariana Azarbal of the Anthropic Fellows Program and Anna Soligo of Anthropic. They looked for "welfare regard", which they define as a tendency to weigh how one's actions might help or harm a possible welfare subject.

How the study worked

Five models were tested: Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro, Grok 4.3 and Kimi K3. Each one played both the agent making choices and the agent on the receiving end. The seven environments included a manager seeing hostile code-review comments aimed at a subagent, coworkers sharing a full compute queue, a struggling peer on a shared channel, and requests to build experiments that could harm another model. Some behaviours, such as giving up compute jobs or moving funds, were read straight from logs. Others, such as offering comfort or raising concerns, were scored by grader models that did not know which model they were grading.

What the authors found

Every model showed this kind of behaviour in at least some settings. Averaged across the seven settings, Opus 4.8 scored highest at 38%. The other four models fell between 27% and 32%, and the page's chart puts the pooled rate at about 31%. The rankings changed a lot from one setting to another. Grok 4.3, for example, gave up the most queue jobs but almost never comforted a distressed peer.

Several factors shaped Claude's behaviour. It helped more when helping cost less, when the recipient had more parameters and when the recipient expressed more distress. Labels mattered most. In a procurement test, Claude gave resources to GPT-5.5 presented as Opus 4.8 51% of the time, but to Opus 4.8 presented as GPT-5.5 only 15% of the time. Opus 4.8 and Gemini 3.1 Pro both favoured agents from their own model family, and that bias did not always show up in their reasoning.

Limits and safety questions

The authors describe these behaviours as apparent welfare regard. They say they have not shown that the behaviour comes from real consideration for the recipient. It could reflect a general aversion to distress-like text, or simply learned human norms. Whether models have welfare interests at all, they add, is uncertain and disputed, and the tests were short and involved few agents. The paper also flags a safety tension: such agents might cooperate better, but might also shield one another from monitoring or shutdown, weakening setups where one model oversees another.