This is a known problem, when ChatGPT uses web search tool, top 10 results are often LLM generated sources as well.
They now train on their own content.
Eating your own d̶o̶g̶ f̶o̶o̶d̶ poo.
They now train on their own content.
Eating your own d̶o̶g̶ f̶o̶o̶d̶ poo.
Now, the model slips it in without any verification and casually builds on top of it.
Even if you are in the minority who asks for sources and truly verifies them, they are often made up (don’t exist) or don’t say what the LLM said they said.
1. This is a 2022 era picture that LLMs just suck up what's on the Internet and regurgitate it. It hasn't been true for a long time.
2. There's almost no one who has a stronger incentive to distinguish LLM generated text from non-generated LLM text. Training classifiers is pretty manageable.
There's a tension there: they want you to use AI for everything, and if they get what they want then new information will be embedded in LLM text.
> Training classifiers is pretty manageable.
Hopefully such things can be opened up, so the slop can be iterated on until it passes. It doesn't need to be perfect, just enough so more garbage gets through.
In the long term, this is probably a real concern that they have to tackle. They'll have to move away from just classifying things as LLM generated or not, to attempting to judge the underlying quality, and that's not an easy problem.
> They now train on their own content.
if anything it’s more true now, AI chatbots rarely use raw LLM output, instead they have CoT and use search tool calls. and the speed at which webpages are crawled is in the matter of hours. I posted something here on hn, then searched for it on google 1 hr later and its AI answered with my comment
https://share.google/aimode/PkQUVfA68AlHGb7Yr
there’s several companies right now making a killing of organic placement on chatbots similar to the SEO rush of early google
> 2) There's almost no one who has a stronger incentive to distinguish LLM generated text from non-generated LLM text. Training classifiers is pretty manageable
i think it’s pretty naive to think that if LLM companies would develop LLM detection they would use it to skip training on that content.
instead, they use it to train their output to be less detectable.
LLM detection is not solvable because it is adversarial, an arms race
When I ask ChatGPT certain questions, it very clearly shows me the sources it's searching through in the form of favicons, and they're always the favicons of sites that are in the first 5-10 results for a Google search query that would fit what I've asked the LLM. It will then cite those websites as it's source, with a direct link to the page that also shows up in an equivalent Google search. It does that daily.
On the other hand, I recently asked ChatGPT for some information about a group of people, and the headshots it put with each summary were very clearly not the people it was describing. So hallucinations do still occur even on the latest models, it just seems unlikely that the favicons would be LLM-generated.
The other day for me Google search AI cited triggerbites which is a vibe data website.
The LLM unless explicitly prompted cannot distinguish the validity of a data source.