Nearly 5 per cent of new Wikipedia pages that are published in English seem to contain text generated by artificial intelligence, which could reduce the site’s reliability.
at Princeton University and his colleagues were wondering about the implications of recently rolled-out AIÂ systems called large language models and how these may affect sources of information.
To learn more, they first ran the AI detection tools Binoculars and GPTZero on more than 14,000 Wikipedia pages published in English, French, German or Italian before March 2022, eight months prior to ChatGPT’s release.
Advertisement
The tools flagged fewer than 1 per cent of these pages as having AI-written content. The true figure is probably nearer to 0 per cent, seeing as ChatGPT hadn’t yet been rolled out, says Brooks. “It’s definitely the case that any individual tool will have some false positives.”
Next, the tools were used on nearly 13,000 Wikipedia pages created in August 2024, which suggested that 4.36 per cent of those published in English contained AI-written content. The rates were generally lower for French, German and Italian-language pages.
“To me, it isn’t that surprising, given it’s happening pretty much all over society,” says team member , also at Princeton University. “Wikipedia is particularly important [as a source of knowledge] just because it’s being used in every single large language model as training data.”
The researchers didn’t identify individual pages that they believed had AI-generated content, but Brooks says a number of incidences of its use were by small business owners who created pages about themselves and their companies.
They also believe that the owner of a manor house in England who was looking to sell their property made a Wikipedia page to amplify its importance. Other examples of AI’s use were politically motivated or to translate pages from one language to another, says Brooks.
Peskoff says the use of AI on Wikipedia could be problematic if left unchecked, primarily because of its unreliability. “This is something that can scale up very, very quickly,” he says. “Right now, it’s 5 per cent, but you don’t know where it’ll be in two years.”
at Staffordshire University in the UK says that the tools the researchers used aren’t infallible, but they are the best we have for detecting this sort of content.
Wikipedia, which is monitored by editors, will probably get better at removing AI-generated content, she says.
“I suspect actually that Wikipedia will become one of the few bastions of human-generated information, unlike other semi-moderated spaces that are rapidly filling up with AI-generated information,” she says. “The particular culture of Wikipedia’s editor-moderators is well suited to put up a battle against generative AI content, and I salute them for it.”
A spokesperson for the Wikimedia Foundation, which owns Wikipedia, says that the paper has some limitations, such as the fact that the tools used weren’t trained to detect AI-generated content on Wikipedia specifically, and AI policies vary on Wikipedia sites around the world.
“The English Wikipedia community has not forbidden the use of AI as long as it’s properly verified and meets Wikipedia’s content policies,” says the spokesperson. “Other language communities considered in the study may have made other choices and those choices will affect the interpretation of the results.”
arXiv
