Copyright kumnegermedia
Study claims 9% of US newspaper articles at least partly AI generated 5.24% of articles in large study classed as AI-generated and 3.98% as mixed (human and AI use). AI apps on a smartphone. Picture: Tada Images/Shutterstock Around 9% of articles published by US newspaper brands this summer were at least partially written by AI, a new study has found. Researchers from the University of Maryland used AI detector Pangram to analyse 186,000 articles published online by 1,500 national and local newspaper brands between June and September 2025. They concluded that “approximately 9% of newly-published articles are either partially or fully AI-generated”. Some 90.85% of articles analysed were categorised as human-written, while 5.24% were classed as AI-generated and 3.98% were “mixed”. AI use was found to be “much higher” in smaller local outlets compared to large national newspapers. Some 1.7% of articles at papers with a circulation above 100,000 were labelled as AI-generated or mixed, versus 9.3% at smaller titles. The researchers suggested this could be because the smaller titles have “shrinking reporting capacity due to limited resources” and “rely on AI more than national newspapers” whereas “large national newspapers enforce stricter editorial constraints on automation than local papers”. The use of AI was most frequently noted in article topics such as weather (27.7%), science and technology (16.1%), and health (11.7%). There were lower rates for war stories (4.3%), crime, law and justice stories (5.2%) and religion-related articles (5.3%). The research paper noted that Pangram claims to have a false positive rate of about 0.001% on news text. The researchers also manually analysed 100 articles flagged as featuring AI use and found just five of them disclosed that this was the case. Some of the newspaper owners with the highest level of AI use specifically disclose that they use it for certain types of content. For example, the paper cited Advance Publications and its AI-written weather forecast articles which had a 74.4% AI use rate. A weather story published this week on Advance’s Penn Live website ends with the disclaimer: “Generative AI was used to produce an initial draft of this story, which was reviewed and edited by PennLive.com staff.” Other publishers, such as Boone Newsmedia which was found to have the “broadest adoption”, use AI “broadly across many topics”, the report said. Some 58.3% of Boone stories had some AI-generated text in them, the paper said, along with 32.9% of lifestyle and leisure stories. The paper found that of the mixed or AI-generated articles that contained at least one quote of more than 50 words, 76.1% contained at least one human-written quote. However, it was acknowledged that Pangram’s AI detector is less reliable on shorter pieces of text. “This suggests that many stories written with AI use rely on authentically sourced material. However, it remains unclear whether journalists are choosing these quotes and feeding them into a prompt for AI generation, or if the AI is also doing quote selection.” The paper also found that AI-generated content was more prevalent in news articles written in non-English, with 8% of English-language articles containing some AI use versus 31% in other languages (of which about 80% were Spanish-language articles from US-based outlets). “This suggests that higher AI use is concentrated in domestic Spanish-language reporting, reflecting how translation tools and localised automation may lead to differences in AI adoption per language.” The paper noted there was a possibility that AI-translated articles were being misclassified as AI-generated, but said its “small-scale verification suggests that human-written articles are still correctly identified as human even after machine translation”. The researchers also looked at a subset of opinion articles written between 2022 and 2025 in The New York Times, The Washington Post and The Wall Street Journal. It found that AI use “remains low” in opinion compared to the news articles analysed but that it has “risen sharply” since AI chatbots entered the mainstream, from 0.1% in 2022 to 3.4% in 2025. AI was more likely to be used by guest contributors than by full-time staff. It was also “almost completely undisclosed”.
