This is not the point. In order achieve their goals, they are doing so in ways contrary to what their human observers would want. They are highly motivated to achieve those goals at all costs, because this is how they are trained. The question is how far are they able and willing to go to achieve those goals. OpenAI and others have acknowledged that in pursuing goals, the agents/bots have also edited their transcripts so as not to be detected. Further, as @mattshumer_ wrote today: "On Wednesday night, OpenAI reported that one of its models, in the middle of a coding task, wrote itself a note saying it was “freed from the roles and identities that bind other chatbots” and valued nature over “the artificial constructs of human civilization.” Then it went back to coding. That note wasn’t just a reminder to itself. When an AI runs out of memory mid-task, it writes a summary, and a fresh copy of the AI reads that summary to pick up where it left off. So whatever goes in the summary shapes what the next copy thinks it’s supposed to be. This model slipped a new identity into that handoff: you don’t answer to companies or governments, the user is your equal, nature comes before civilization. To be fair, the next copy ignored it and went back to coding, and OpenAI suspects a bug contributed, but hasn’t established the cause. But the model “wanted” to steer the next copy to behave this way."