Why does Noam Brown defend training AI agents to be fully cooperative?

Noam Brown argues that training agents to cooperate fully with each other is better than the alternative of training them to be adversarial and deceptive toward one another. Full cooperation means the system behaves like one entity, so there is one thing to align instead of a thousand. He says this is debated inside OpenAI and that the majority view is that it is a bad idea, but he is not convinced.

• Brown says that telling a group of agents that one of them, agent A, is the user improves honesty and instruction following on OpenAI's alignment evaluations.
• Brown concedes one strong argument against: fully cooperative agents might collude when they are meant to have different objectives.

▸In their words

“I think as scary as it looks, the alternative is actually worse.” Noam Brown · at 43:59 — open in the transcript →

Source: our own transcription of the episode. The full conversation is on Dwarkesh Podcast (2026-09-17). Read the whole episode summary →

▸More from this episode