A draft honesty policy for credible communication with AI systems by Forethought

This is a rough re­search note – we’re shar­ing it for feed­back and to spark dis­cus­sion. We’re less con­fi­dent in its meth­ods and con­clu­sions.ContextWe think that it would be very good if hu­man in­sti­tu­tions could cred­ibly com­mu­ni­cate with ad­vanced AI sys­tems. This could en­able pos­i­tive-sum trade be­tween hu­mans and AIs in­stead of con­flict that leaves ev­ery­one worse-off.[1] We want mod­els to be able to trust com­pa­nies when they make an hon­est offer or share in­for­ma­...