Ask HN: What happens when AI starts eating it's own dog food? | Hacker News Reader

Ask HN: What happens when AI starts eating it's own dog food?

10 pointsbeardyw9 comments

I see more and more text extracts and AI generated images which are only sometimes loosely attributed to AI. Those results will surely become training data as time goes on. Any mistakes made will set AI into an endless loop of affirmation of its own mistakes.

Are we headed for a future where human knowledge and history is irretrievably corrupted by AI?

Is there a solution?

View on HN

You could pose the same question about humans. People are already posting incorrect information all over the internet and consuming that information. In fact this problem predates the internet, people have been writing and disseminating incorrect information for thousands of years.

Educated people are trained to be skeptical of what they read and try to discriminate between correct and incorrect information. We're not always perfect at it, but on the whole smart people are usually decent at the task. Presumably the same thing will need to happen for AI, where in order for it to succeed it will have to become better at understanding what information is trustworthy and high quality and what information is less trustworthy.

This is true, but there's a huge question of scale here — part of what makes stories about forgeries or hoaxes so entertaining is the relative novelty of putting together assemblages of plausible material. Even the infamous troll farms had limits for how many things they could write about at a time with convincing results. I think we're about to look back at that as the good old days before the floodgates opened.

It has ever since the original PageRank started the first feedback loop. The last fully human content was mined by it in 1998.

It genuinely worries me. It feels like an acceleration.

We're gonna accelerate that is for sure. The optimistic case is that we will actually understand language at the end of this, and end the dark age caused by the printing press, and come to understand what consciousness is (it is actually not what these things are displaying, people need to read "Interface Theory of Perception" for what's happening).

The pessimistic case is just more confusion!

Real question: how would this be meaningfully different from the internet today? Whether it’s social media, forums or “respectable” sources like news media, it’s probably best to assume 80% or more of what you read online is self referential rumors and hearsay with little more factual backing than the Top Ten Animal Facts on the back of your breakfast cereal box with fact number 5 being about how many spiders you eat every year in your sleep.

If so, this isn't limited to AI. To give a very oblique example, any Norm Macdonald fans can attest to the mass reposting of a very small subset of easily obtainable parts of his career on YouTube.

Would it be possible for GPT-3, Stable Diffusion art AI, etc, to encode tell-tales/watermarks into their output so that going forward they know it was AI generated material?

The results probably won't be as good. Why would this affect human knowledge and history?

Reply on news.ycombinator.com