Scott Cory

I would gently push back on the underlying premise that if the government agrees to a usage policy restriction, that's ironclad, but if it's just a law or policy, that's no guarantee at all. Why would Anthropic think that their earlier usage policy forbidding surveillance was sufficient to guarantee their models could not be used for this? My main argument is that usage policies are only one part of a layered set of safeguards. Here's how I think about this: 1. The safety stack travels with the model. The Department was not asking us to modify how our models behave. Their position was, build the model however you want, refuse whatever requests you want, just don't try to govern our operational decisions through usage policies. For whatever risk surface area remains, our safety stack, refusal policies, and guardrails become another protection. And those technical controls are often more reliable than contract clauses anyway. Our contract gives us control over the models and safety stack we deploy, and the ability to improve them over time. 2. AI experts directly involved. Instead of hoping contract language will be enough, our contract allows us to embed forward deployed engineers, commits to giving us visibility into how models are being used, and we have the ability to iterate on safety safeguards over time. If our team sees that our models aren't refusing queries they should, or there's more operational risk than we expected, our contract allows us to make modifications at our discretion. This gives us far more influence over outcomes (and insight into possible abuse) than a static contract provision ever could. 3. U.S. law already constrains the worst outcomes. We accepted the “all lawful uses” language proposed by the Department, but required them to define the laws that constrained them on surveillance and autonomy directly in the contract.  And because laws can change, having this codified in the contract protects against changes in law or policy that we can’t anticipate.