Making credible deals with AI — how labs can trade with scheming systems
If a weakly superhuman AI schemes toward misalignment, one risk-reduction strategy is to offer it things it values (donations to aligned causes) in exchange for useful behavior (revealing misalignment, better auditing techniques).
The bottleneck is credibility: a scheming AI knows the lab controls its training data, tool outputs, and runtime context, so why trust any deal is real rather than a fabricated test or honeypot?
LessWrong post proposes mechanisms (foundation websites, honesty strings, signed contracts) designed to make such deals cryptographically or structurally credible to the AI.