The rapid growth of AI models from OpenAI and its competitors has flooded the internet with low-quality information that enters the training datasets of new AI models and causes errors.
The rapid growth of AI models from OpenAI and its competitors has flooded the internet with low-quality information that enters the training datasets of new AI models and causes errors.
This is reported by Futurism, citing The Register, which compares the situation to the demand for «low-background steel,» which was produced before the detonation of the first nuclear bombs, starting in July 1945. These explosions released radionuclides and other particles that seeped into virtually all steel produced thereafter.
This makes modern metals unsuitable for use in some highly sensitive scientific and medical equipment. A significant source of low-background steel even today is the battleships of World Wars I and II, including the huge naval fleet that was sunk in 1919.
«This allowed us to have an almost infinite supply of low-background steel. If it weren’t for that, we would be stuck. But if you collect data until 2022, you can be pretty sure that it has minimal, if any, contamination from generative AI. Everything before that date is ‘safe, good, clean,’ everything after that is ‘dirty,’» the scientist noted.
In 2024, Chiodo co-authored a paper arguing that there should be a source of «clean» data, not only to prevent model collapse but also to ensure fair competition among AI developers. He argues that early adopters have an advantage because they are the only ones who benefit from a cleaner source of training data before AI pollutes the internet.
Scientists currently disagree on whether AI models collapse due to data pollution, but many researchers have been sounding the alarm for years.
«It’s unclear right now to what extent model collapse will be a problem, but if it is a problem, and we have polluted this data environment, cleanup will be extremely expensive, probably impossible», — says Chiodo.
There are already issues with «augmented search generation,» which AI models use to supplement their old training data with information gleaned from the internet in real time. But this new data is not guaranteed to be free of AI fakes, and some studies have shown that it leads to chatbots giving far more «dangerous» answers.
After OpenAI and others reported diminishing returns from their latest models in late 2024, some experts said scaling had hit a ceiling. And if that data becomes increasingly laden with «information noise,» that hurdle will become even more insurmountable.
Chiodo suggests that stricter rules, such as labeling AI content, could help label some of this «noise,» but this will be difficult to implement, especially given companies’ resistance to government regulation of the AI space.
