Beyond half of the articles available online would be written by AI. This was revealed by a recent study conducted by the SEO analysis company Graphite who recently published the results of a research during which it emerged that a November 2024 the amount of AI-generated articles published online has surpassed that of texts written by humans. It is the first time something similar has happened: for a short period, machines produced more textual content than “flesh and blood” authors. The study, which examined 65,000 articles in English collected from the vast open source archive Common Crawlhowever, also showed that the growth of artificially generated content has stabilized since May 2024.
A key element of this research is that, despite the amount of text written by algorithms, most of it does not appear in search engines or in chatbots like ChatGPT: according to Graphite, more than 80% of the results shown by Google and conversational assistants still come from articles written by people. The data suggests that, at least for now, artificial intelligence has not “taken control” of online information, but rather has alongside man in a form of co-production that is increasingly widespread and difficult to distinguish.
Graphite’s investigation and the method used
Graphite’s investigation was based on a methodology very precise. To assess the spread of AI-created content, the researchers used an AI detector developed by SurferSEOcapable of analyzing 500 word text blocks and estimate the probability that they were written by a language model. An article is considered “artificial” if over 50% of its content was generated by an AI system. Before applying this classification, the scholars wanted to verify the precision of the tool, calculating the rate of false positives (i.e. human texts incorrectly labeled as artificial) and false negatives (texts generated by AI not recognized as such). The analysis showed a false positive rate of 4.2%testing articles published before the launch of ChatGPT, a period in which, in all likelihood, the texts were entirely written by people. The false negative rateinstead, stopped at 0.6%using articles produced by GPT-4o, one of the most advanced linguistic models available at the time of the search, using a very detailed and defined prompt to produce texts on various topics. From these preliminary checks it emerged that the classification made with the aid of SurferSEO detector turns out to be quite reliable with a AI correct detection rate of 99.4%.
The results collected outlined a clear trend: from November 2022the moment ChatGPT went public, automated content production literally exploded. Alone 12 months AI-generated articles have moved on to represent approximately the 39% of published texts, to the point of briefly surpassing human ones a November 2024. However, since 2025 the situation has stabilized and the two values have aligned. Graphite hypothesizes that this slowdown is due to the poor performance of texts produced by AI in search engines: Google’s search engine, for example, tends to give priority to content written by people, considered more reliable and relevant. A second research conducted by the same company found that 86% of articles in Google Search are of human originwhile only the 14% comes from artificial intelligence systems. A similar data surprisingly also emerges from the main AI chatbots: ChatGPT and Perplexity cite for82% articles written by humans and only for the 18% articles whose text is generated using artificial intelligence.
The limitations of the study
The Graphite study that discovered the percentage of diffusion of AI texts online, as interesting as it is, still presents some limitsclearly highlighted also by the authors of the study themselves. First of all, it should be remembered that the Common Crawl databasethe data source used for the research, It doesn’t cover the entire web. Many paid sites prevent their content from being indexed, which therefore is not included in the samples analyzed in the study. This means that the percentage of articles written by humans may actually be higher than estimated. Besides this, the authors of the study admit that it can be difficult to distinguish articles written by AI and then edited by humansin addition to the fact that AI models continue to improve and, therefore, may become increasingly difficult to detect.
