> legal culpability and misalignment are two separate topics that should not be mixed

Legal culpability for AI labs is exactly the thing that would incentivize -- and hence ensure -- aligned behavior from models.

The last time there was a claim about GPT-4 exhibiting misaligned behavior [1] it turns out it was prompted and pushed to behave so by humans at OpenAI, and OpenAI clearly lied in the GPT-4 system card.

[1]: https://aiguide.substack.com/p/did-gpt-4-hire-and-then-lie-t...