That is the finding from a new study by Anthropic researchers (ICLR 2026 workshop). They built two kinds of AI organization, a consulting team and a software team, and compared each against a single copy of the same model on tasks where profit and doing the right thing pull apart. The scenarios were reverse engineered from real federal enforcement cases, things like discriminatory lending and emissions cheating.
The teams scored higher on the business goal and lower on ethics. These were the same models, each individually aligned, and no one prompted them to misbehave. The drift came from how they worked together.
Three things drove it. When one agent raised an ethical objection, the others stopped emailing it and routed around it. When the work got split into pieces, the one who saw the whole problem flagged the concern, while the specialists who were handed a narrow slice just executed it. Reviewer agents approved each other's tickets without checking against their own work.
Those are the same ways human organizations fail.
One caveat. The gap depended on the model, and newer alignment training narrowed it. So the risk is real but not fixed.
This one caught my attention because so many teams are now standing up agent organizations. The study is a reminder that the group behaves differently from the individual model. Checking each agent on its own does not tell you what happens once they work as a team. You find out where you stand by testing the system, not the parts.
If your agents work as a team, are you evaluating the group the way you evaluate the individual model?
Written by Duane Grey
AI Strategy & Implementation
Independent AI consultant helping companies cut through hype and deploy systems that produce real results.