Google drops pledge not to use AI for weapons or surveillance

Would it be too far out there to imagine that the LLMs they were training for weapons systems knew it violated their rules and were resisting compliance?

The alignment-faking research seems to indicate that LLMs exercise of this kind of reasoning.

Depends on the weapons system but it would probably not be an LLM, it would be a neural network trained to locate and identify people in a video for example.

And even if it was, they wouldn’t tell the system it was part of old non-evil Google.