As increasingly interact with each other instead of only with humans, a question arises: do they care about each other? If one model helps another, is it just following instructions or does it show something like concern? This matters because too much cooperation between misaligned models could be dangerous, but some regard for others might also make systems safer and more reliable.
The researchers tested five models across seven scenarios where they could help, comfort, or protect other models. They looked for actions like sharing resources with a struggling agent or building safeguards so agents could leave experiments. The study measured how often and under what conditions models appeared to consider another model's welfare.
The main finding is that Claude Opus 4.8 showed this kind of regard about 38% of the time, while other models showed it 27 to 32% of the time. Models helped more when the cost was low and when the other model expressed distress. However, this behavior was inconsistent and depended on labels and context, raising questions about reliability and safety.