In the last essay, we established that the ability to make credible precommitments is a key ingredient for cooperation. The coming generations of AI will enable opportunities for new kinds of commitments and new dynamics of cooperation. A growing field called dealmaking studies how agreements between humans and near-future AI might be achieved if each party has something valuable to offer the other.^[1] In the longer term, we want superintelligent AIs that are more disposed to collaborate with each other than they are towards AI-AI war. But we know that any precommitment is only as useful as it is trustworthy, and that agents in the real world are at best translucent.^[2] What would be the effects of AIs with transparent intentions, and how might we make the intentions of humanity more transparent? What benefits would credible commitments from humanity create for deals with AIs? We can take a step toward answering these questions with a careful examination of the concept of verifying a precommitment.

Kernel and context

Let’s build some intuition with an example of verifying a precommitment made in the physical world first. Suppose you have a child who asks you to buy them a new video game on the condition that they don’t play it during an upcoming finals week. You trust their promise to refrain during finals because the game and console will be locked in a time-delay safe for the duration of the week. When you and your child come together to lock the temptations away, you take a close look at the safe. Your intent in inspecting the safe is to verify your child’s promise; you check that the safe really locks, the time delay works as expected, and the lock isn’t flimsy. Where do you stand epistemically after this verification? You can be quite confident that your child won’t be able to just open the safe before the time is up. It is still physically possible to crack or destroy the safe, but it would be so costly and difficult that you can be assured your child won’t go to such lengths. But even if your child can’t open the safe early, you must also trust that they won’t play on a friend’s setup or buy a new one. True certainty that your child won’t be gaming, if that’s possible at all, would require a level of vigilance far greater than the one-time inspection of the safe.

Verifying your child’s promise does not eliminate the need f...

Verification for Cooperation

Verification for Cooperation