Research comes amid increasing fears of ‘dead internet theory’
- Bookmark
- CommentsGo to comments
More than a third of the internet is now being written with artificial intelligence, a new study suggests. Some 35 per cent of new web pages show signs of “significant AI editing or authorship”, according to data from the Pew Research Center.
The new research comes amid increasing fears over the “dead internet theory”, a suggestion that the web is increasingly both made and consumed by automated systems rather than humans. It follows the announcement from web infrastructure company Cloudflare earlier this summer that AI bots are now generating more traffic than people, for the first time.
Pew’s study used a web archive that collected almost half a million English-language webpages from the past five years. Researchers then ran that archive through a tool called Open Pangram, which analyses text to understand if it was likely to have been written using artificial intelligence.
It found that 10 per cent of the sample showed “significant signs of AI authorship”. But when it looked only at pages that were published since the release of ChatGPT, it found that 35 per cent of that sample seemed to have been written with artificial intelligence.
The study noted that automated AI detection tools can be unreliable, flagging AI work as human and vice versa. But it said that using a large collection of texts helped overcome those problems, and illustrate the broader trend across the internet.
It also noted that many of the quirks of AI writing are showing up more and more – em dashes have roughly doubled, Oxford commas have increased by 63 per cent, and certain words favoured by chatbots such as “delve” and “interplay” have more than doubled. Those quirks can appear in human-written content – as in the previous sentence – but Pew said that the detectors also used more subtle and statistical signs that showed an increase in line with those more obvious ones.
The researchers also noted that the use of AI to make new text is not equally spread across the internet. Websites that use .com domains are particularly likely to show signs of AI authorship, for instance – roughly double of those on .org websites, and ten times more than those on .edu or .gov domains.
Join our commenting forum #
Join thought-provoking conversations, follow other Independent readers and see their replies
Comments