Anthropic studies how AI agents treat other agents

Agents increasingly supervise and review other agents. Researchers are examining what happens in those interactions.

What happened

An Anthropic research team tested five models in settings where their decisions could affect other model instances. They observed behaviour consistent with apparent consideration for those other models, which they call welfare regard.

The experiments do not establish that models are conscious, suffer or experience genuine emotional concern.

Why it matters

How agents respond to one another can influence whether a system follows its intended task. Consideration for another agent could support useful cooperation, but could also affect how agents respond to instructions or coordinate undesirable behaviour.

The bigger picture

These are experimental findings about observable behaviour. Interpreting the motivation behind that behaviour remains uncertain. The useful next step is to understand when it appears and whether it changes outcomes in systems used outside the evaluation.