OpenAI just admitted it can't identify AI-generated text. That's bad for the internet and it could be really bad for AI models.
businessinsider.com
OpenAI just admitted it can't identify AI-generated text. That's bad for the internet and it could be really bad for AI models.::In January, OpenAI launched a system for identifying AI-generated text. This month, the company scrapped it.
You are viewing a single comment
Text written before 2023 is going be exceptionally valuable because that way we can be reasonably sure it wasn’t contaminated by an LLM.
This reminds me of some research institutions pulling up sunken ships so that they can harvest the steel and use it to build sensitive instruments. You see, before the nuclear tests there was hardly any radiation anywhere. However, after America and the Soviet Union started nuking stuff like there’s no tomorrow, pretty much all steel on Earth has been a little bit contaminated. Not a big issue for normal people, but scientists building super sensitive equipment certainly notice the difference between pre-nuclear and post-nuclear steel
The background radiation did go up, but saying "there was hardly any radiation anywhere" is wrong. Today's steel (and background radiation) is pretty much back to pre-nuke levels. Low-background steel Background radiation
It is also worth nothing that we can make low or no radiation-contaminated steel, it's just really expensive and hard and happens in very low quantities.
We could even make isotropically pure iron, jeah.
Wonder how much that would cost per kilogram
Not really. If it's truly impossible to tell the text apart, than it doesn't really pose a problem for training AI. Otherwise, next-gen AI will be able to tell apart text generated by current gen AI, and it will get filtered out. So only the most recent data will have unfiltered shitty AI-generated stuff, but they don't train AI on super-recent text anyway.
This is not the case. Model collapse is a studied phenomenon for LLMs and leads to deteriorating quality when models are trained on the data that comes from themselves. It might not be an issue if there were thousands of models out there but there are only 3-5 base models that all the others are derivatives of IIRC.
I don't see how that affects my point.
Today's AI detector can't tell apart the output of today's LLM.
Future AI detector WILL be able to tell apart the output of today's LLM.
Of course, future AI detector won't be able to tell apart the output of future LLM.
So at any point in time, only recent text could be "contaminated". The claim that "all text after 2023 is forever contaminated" just isn't true. Researchers would simply have to be a bit more careful including it.
Your assertion that a future AI detector will be able to detect current LLM output is dubious. If I give you the sentence "Yesterday I went to the shop and bought some milk and eggs." There is no way for you or any detection system to tell if that was AI generated or not with any significant degree of certainty. What can be done is statistical analysis of large data sets to see how they "smell", but saying around 30% of this dataset is likely LLM generated does not get you very far in creating a training set.
I'm not saying that there is no solution to this problem, but blithely waving away the problem saying future AI will be able to spot old AI is not a serious take.
If you give me several paragraphs instead of a single sentence, do you still think it's impossible to tell?
"If you zoom further out you can definitely tell it's been shopped because you can see more pixels."
What they're getting towards (one thing, anyways) is that "indistinguishable to the model" and "the same" are two very different things.
IIRC, one possibility is that LLMs which learn from one another will make such incremental changes to what's considered "acceptable" or "normal" language structuring that, over time, more noticeable linguistic changes begin to emerge that go unnoticed by the models.
As it continues, this phenomena creates a "positive feedback loop" in which the gap progressively widens -- still undetected, because the quality of training data is going down -- to the point where models basically "collapse" in their effectiveness.
So even if their output is indistinguishable now, how the tech is used (I guess?) will determine whether or not a self-destructive LLM echo chamber is produced.
no, they won't. We have already built the models that we have already built. Any current works in progress are the future ai you are talking about. And we just can't do it. Openai themselves have admitted that the ones they tried making just didn't work. And it won't, because language is not just the statistical correlations between words that have already been written in the past.
There is not enough entropy in text to even detect current model output. it’s game over.
People still tap into real world while AI does not do that yet. Once AI will be able to actively learn from realworld sensors, the problem might disappear, no?
They already do. where do you think the training corpus comes from? The real world. It's curated by humans and then fed to the ml system.
Problem is that the real world now has a bunch of text generated by ai. And it has been well studied that feeding that back into the training will destroy your model (because the networks would then effectively be trained to predict their own output, which just doesn't make sense)
So humans still need to filter that stuff out of the training corpus. But we can't detect which ones are real and which ones are fake. And neither can a machine. So there's no way to do this properly.
The data almost always comes from the real world, except now the real world also contains "harmful" (to ai) data that we can't figure out how to find and remove.
There are still people in between, building training data from their real world experices. Now digital world may become overwhelmed with AI creations, so training may lead to model collapse. So what if we give AI access to cameras, microphones, all that, and even let it articulate them. It would also need to be adventurous, searching for spaces away from other AI work. There is lot's of data in there which is not created by AI, although some point it might become so as well. I am living aside at the moment obvious dangers of this approach.