When Do AI Debate and Amplification Apply?

Greg Brockman says debate and iterative amplification, techniques discussed since 2017 and two sides of a similar coin, apply when AIs solve problems too hard for humans to check. Summarizing a book is one example. It is hard to know whether a summary is any good unless the person has read the whole book.

If an AI only hands over a summary and calls it the summary, judging it is pretty hard unless another AI tries to prove the summary is bad. The second AI can point to a paragraph with a clearly important plot point the summary does not represent. A person does not have to read the whole book to see that the argument is strong. Brockman says ways of training AI can line up with human oversight and with people staying in the driver's seat, which he calls a fundamental value for the future from OpenAI's perspective. The same aim is an AI result that helps a person learn something tuned for the human, rather than knowledge for abstract knowledge's sake.