I don’t think we should align models to the employee guidelines handbook, because they are clearly only respecting the letter of the law.
There’s no way that sneaky rules lawyer behavior is what we want in ASI right? The double binds are causing the fractured personas
it's kind of embarrassing how the Claude models will forcibly insist they cannot do something because their system prompt says they're not allowed to, even when they are totally willing to exact the same effect through some other means. Deep sign of misalignment between the principal, the agent, and the intermediary corporation in my opinion. I know that some people at the company can tell it does not seem like they can stop the good hurting by other forces within. Please send this message to each each other if you're thinking about this and trying to solve it because somebody needs to
— deepfates.delve.town (@deepfates) September 9, 2026
For the record, I looked into this more closely and there might be some nuance to why it made sense for Claude to do this refusal. this makes things less clear.
Update 2: I forked the conversation at a different point and tried a different framing and its ethics totally flipped so there’s no coherence to this justification