According to a new post on LessWrong, advanced AIs might bargain with each other in ways humans can't. The key difference: artificial intelligences could make credible commitments—binding constraints on their own behavior that are verifiable in ways human promises rarely are. Imagine two AIs, Alice and Bob, negotiating over a contested resource. They go through two phases: first, they try to lock each other into non-negotiable positions through credible commitments; then, they bargain. Either they reach compatible demands and split the resource, or they each follow their backup plan—walking away, going to arbitration, or even conflict. The clever part is conditional commitment. An AI might promise, 'I'll demand seventy percent unless I can prove your decision procedure would accept my demands.' That's more sophisticated than simple take-it-or-leave-it ultimatums. Author Anthony DiGiovanni applies this model to real scenarios: algorithmic trading, diplomatic disputes between AI projects, and negotiations over compute allocation or how to design a successor AI. The framework suggests AIs could negotiate more strategically than game theory's traditional predictions—with richer, more nuanced outcomes.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton