I agree that reaching a pre-deployment treaty is the best option (indeed, this seems outright good rather than just the least bad to me)
You touched on this in the conclusion, but I feel it is a sufficiently important point to foreground more: I think the threat (perhaps implicit) of violence is a key element of successful negotiation and agreement-building. You didn’t much discuss why the deployer state would agree to be constrained by an agreement in section 2, but it seems to me the strongest reason for them to do so would be the credible threat of being invaded if they do not. I think such a threat would be credible, as even if the deployer state is the pre-eminent global power at that time, as seems fairly likely, the harm to other nations’ interests of idle acquiescence could be immense. Thus, rival countries would plausibly be highly motivated to band together to fight a rogue deployer nation before the AI is deployed and (putatively) invasion becomes impossible because the deployer has such an impressive lead in military and other technology.
I think given this credible threat, the deployer would likely be willing to negotiate in good faith, in which case your credible commitment issue comes to the fore.
There was an incomplete sentence at the end of 3.1 “The Biden Administration and allies ”
I agree that reaching a pre-deployment treaty is the best option (indeed, this seems outright good rather than just the least bad to me)
You touched on this in the conclusion, but I feel it is a sufficiently important point to foreground more: I think the threat (perhaps implicit) of violence is a key element of successful negotiation and agreement-building. You didn’t much discuss why the deployer state would agree to be constrained by an agreement in section 2, but it seems to me the strongest reason for them to do so would be the credible threat of being invaded if they do not. I think such a threat would be credible, as even if the deployer state is the pre-eminent global power at that time, as seems fairly likely, the harm to other nations’ interests of idle acquiescence could be immense. Thus, rival countries would plausibly be highly motivated to band together to fight a rogue deployer nation before the AI is deployed and (putatively) invasion becomes impossible because the deployer has such an impressive lead in military and other technology.
I think given this credible threat, the deployer would likely be willing to negotiate in good faith, in which case your credible commitment issue comes to the fore.
There was an incomplete sentence at the end of 3.1 “The Biden Administration and allies ”