Micah Carroll

In the claim ledger

4 promoted claims about them. Assessments are the model’s knowledge, not verification.

• Existing definitions of AI manipulation fail on either implementability or generality.

  plausible · Characterizing Manipulation from AI Systems

• A person's initial state is a bad counterfactual baseline because natural change is normal and often good.

  consistent · Characterizing Manipulation from AI Systems

• Deciding which influences count as harmful is an irreducibly political judgement.

  consistent · Characterizing Manipulation from AI Systems

• The proposed DR-MDP objectives are, with exceptions, tractable with existing RL methods.

  plausible · AI Alignment with Changing and Influenceable Reward Functions