You've probably read something online recently and wondered: was this written by AI? There's a good chance it was.
New research from the Pew Research Centre, in the US, examined hundreds of thousands of web pages to measure how often they showed signs of AI generation or editing.
Our chart of the week shows that web pages published after ChatGPT was released in November 2022 exhibit significant AI involvement, at 35 per cent, or more than one in three.
This figure drops to 9.6 per cent when pages created before the release of ChatGPT are included in the sample for that period, revealing how much AI has become a feature of content generation.
Before ChatGPT
Our second chart shows that before ChatGPT's release, signs of AI involvement were virtually non-existent at around 1 per cent. But from late 2022, after the launch of ChatGPT, the share began to climb. As Claude, Gemini and other generative AI tools became part of everyday content production, that trend continued.
But the chart also shows something else. It also highlights the gap in uptake of AI-generated content across different online domains. In a sample that traced sites before and after the release of ChatGPT, 9.3 per cent of .com web pages showed significant signs of AI-generated content in 2026, roughly twice the rate of .org pages (4.6 per cent).
The proportion was much lower for .edu and .gov domains, at around 1 per cent. This suggests that many educational and government institutions are choosing not to use AI-generated content on official web pages.
Another study, by Graphite Growth, examined the impact of AI specifically on published articles and suggested that, for the first time, AI has been responsible for writing as many online articles as humans.
What is the internet starting to sound like?
The Pew researchers also assessed patterns in the language created by generative AI, including the words used, sentence structures and punctuation.
The analysis revealed that since 2023, use of em dashes has roughly doubled and Oxford commas are up 63 per cent. Use of words and phrases such as “align with”, “delve”, “interplay”, “showcase” and “testament” – which the researchers have identified as AI-typical – have more than doubled. Alongside this, writing structures such as negative parallelism have nearly tripled. Negative parallelism is a stylistic device which defines something by describing what it is not, as well as what it is, ie “it's not X, it's Y”.
None of these patterns, in isolation, proves that the content was written by AI. But together, the pattern is used to determine the probability of whether AI was involved or not.
There is one important caveat to these studies. AI detection tools are not perfect, and many models have been known to incorrectly classify human-written content as AI-generated, while some AI-generated text is misclassified as human-written.
We should also consider what happens when AI starts influencing human writing. As AI-generated text becomes more common online, could it also start changing the way humans write?
