Would this type of debate be useful amongst humans? For example could a company ask two potential suppliers to use this type of debate against each other to decide which one to buy from?
One obvious problem is that the two suppliers have a lot of private information. E.g. if one supplier claims “my product is made out of baby unicorns”, the other supplier is likely to be unable to refute that, because they don’t actually know the production process of the first supplier. The two suppliers are then incentivized to present the most dishonest versions of their private information.
The AI agents presumably work on the same public information, but computation can generate new information, which can be private. Especially if the agents are allowed a lot of time before the debate to prepare their arguments, Alice could run a costly computation that Bob doesn’t run and such that Bob wouldn’t be able to verify her claims about it during the debate. Then Alice would be free to lie about it.
One obvious problem is that the two suppliers have a lot of private information. E.g. if one supplier claims “my product is made out of baby unicorns”, the other supplier is likely to be unable to refute that, because they don’t actually know the production process of the first supplier. The two suppliers are then incentivized to present the most dishonest versions of their private information.
The AI agents presumably work on the same public information, but computation can generate new information, which can be private. Especially if the agents are allowed a lot of time before the debate to prepare their arguments, Alice could run a costly computation that Bob doesn’t run and such that Bob wouldn’t be able to verify her claims about it during the debate. Then Alice would be free to lie about it.