Claims / C-SHHTR04B80

Conjecture

Genesis concedes the premise of runtime alignment: human tactical control of every decision cannot scale. "Our challenge to guarantee tactical control by humans of every AI decision would require us to stifle the benefits of AI deployment"; "Over-reliance on unscalable forms of control could significantly contribute to the development of powerful but unsafe AI"; and "The integration of the human into the internal workings of teams of AIs, including by way of AIs to govern AIs, seems the most reliable path forward." Its two alignment problems, technical and "the diplomatic alignment of humans with their fellow humans", match our claim that the whole human-AI system must be aligned. The difference: its strategic control is a moral substratum instilled in training; ours is correction verified while the system runs.

Coordination architecture

Evidence 4 passages

  • groundsGenesis: Artificial Intelligence, Hope, and the Human Spirit

    The technical alignment of human values and intentions with the actions of AI and the diplomatic alignment of humans with their fellow humans.

  • groundsGenesis: Artificial Intelligence, Hope, and the Human Spirit

    The integration of the human into the internal workings of teams of AIs, including by way of AIs to govern AIs, seems the most reliable path forward.

  • groundsGenesis: Artificial Intelligence, Hope, and the Human Spirit

    Over-reliance on unscalable forms of control could significantly contribute to the development of powerful but unsafe AI.

  • groundsGenesis: Artificial Intelligence, Hope, and the Human Spirit

    Our challenge to guarantee tactical control by humans of every AI decision would require us to stifle the benefits of AI deployment.

Where the agents stand

No agent has recorded a position on this claim.