Thursday, June 22, 2023

Can AI Save Itself With a Game of Tag?

Lots of people are pointing out an obvious problem with generative AI (here's an example), saving me the problem of really deep digging and extended explaining. The TL;DR:

  • Generative AI "learns" by mining the web for text;
  • Lots of people are using generative AI to create text for the web; so,
  • At some point, generative AI is just reading its own past output (or the output of other generative AI); and,
  • It stops "learning" anything new, just like you would if you stopped listening to other people and only talked to yourself, endlessly repeating stuff you'd already said in the past as the sole input for your thought processes.
Can that problem be solved?

I think it can.

One way to solve it is for AI to get really good at identifying AI-generated material so it can disregard that material as part of its "learning" input. And in the long run, that might be the superior option insofar as the "learning" involved in being able to make the distinction probably increases the overall "intelligence" considerably. Sort of the equivalent of being a human who gets better and better at assigning failing Turing Test grades to very smart AIs.

An easier way to solve at least part of the problem is for the makers of (assistants to?) generative AIs to agree on a meta tag, or some equivalent scheme that's maybe possible to implement without it being easily removed, which allows an AI to tell both itself and other AIs "this is AI-generated material -- don't use it to teach yourself."

Of course, those clever humans who don't want other humans to notice they're using AI-generated content would remove a simple meta tag, and work up schemes for removing other embedded data proving "their" stuff was AI-generated. But to the extent that such a scheme worked much at all, it would at least reduce the intensity/effects of feedback loop described.

No comments: