Comment by aabhay

In my opinion all this discussion of the contract language is a subterfuge. The real question is why the government was requesting this language in the first place. Clearly there’s more to it than a legal battle.

In my mind, the government would be fully happy to use this to surveil citizens (and indeed anyone) with or without any legal basis, but the issue was that Anthropic has a safety stack / training and inference protocols that it follows. Refusals, abuse models, and manual guardrails. They didn’t want to shut those off. Likely there were some very basic technical reasons, some being that the team’s safety posture is fully ingrained in the model itself and thus difficult to remove.

In this document, OpenAI admits that while they are not “turning off” their safety stack, they are completely willing to provide the government with a different model, different guardrails, etc. That should be incredibly concerning. Anthropic was unwilling to do this, cited their ToS, and ultimately had to walk away from the deal. Given that the government (DoW really) framed this in terms of a hilariously stupid position (surveillance and autonomous weapons), Anthropic felt that this was something they could voice to the public and therefore the entire guardrails discussion turned into a “we want the language changed”. Also the government can’t actually compel Anthropic to create new guardrails so they had no choice but to raise the stakes, make this a moral thing, and basically accuse Anthropic of being woke.

IMO this is really sad for OpenAI employees. Yet again Sam Altman proves that he wants to weasel his way around public perception. Folks at the company have to grapple with working for someone of that disposition.