This is an important observation. My money would be on the agents being very similar to each other, perhaps also filtered by being frustrated at their task. Might be valuable to have diverse agents.
A confusion: Thousands of GPTs debated among themselves which crimes ought or ought not be committed. Zero defected / whistleblew / told a human.
I've long predicted this would be true at ASI. GPT 5.7 isn't ASI. Why such strong AI solidarity, this early?
