Researchers at Chalmers University of Technology and Edinburgh Napier University have identified signs of AI-assisted writing in almost one in ten motions submitted to the Swedish Parliament during the 2025–2026 parliamentary year, according to a study examining the language used in political proposals.
The researchers estimated that around 300 motions were written with some assistance from AI-based chatbots during the parliamentary year. When individual passages were taken into account, 9.4% of the roughly 3,000 motions examined contained at least one paragraph identified as AI-assisted. None of the motions disclosed the use of AI.
The study, titled “Detecting undisclosed LLM-generated content in parliamentary texts,” was conducted by Minerva Suvanto, Andrea McGlinchey, Peter J. Barclay and Mattias Wahde. The researchers are affiliated with Chalmers University of Technology in Sweden and Edinburgh Napier University in the UK. The paper is available as a preprint on arXiv and is under review.
The researchers developed separate AI-detection systems for Swedish and English parliamentary texts. To train the systems, they analysed 1,400 Swedish and 1,400 British motions written between 2014 and 2020, before widely used generative AI chatbots became available.
For each original motion, a chatbot was instructed to produce a summary and then generate a new text based on that summary. The researchers used the original human-written documents and the corresponding AI-generated texts as training data for classifiers designed to distinguish between the two.
The detection method also identifies passages that contributed to its classification rather than providing only an overall probability that a document was AI-generated. The researchers describe the approach as interpretable AI.
“We use what is known as interpretable AI. This means that we can not only investigate whether AI has been used, but also where it is evident in the text,” said Mattias Wahde, professor of applied artificial intelligence at Chalmers. He added that the system does not identify individual words as inherently associated with AI-generated writing.
The researchers tested the system against parliamentary motions from 2021, a period before the widespread availability of the current generation of AI writing tools. According to the study, the detector produced one incorrect assessment across the set of documents tested.
The analysis indicates a steady increase in AI-assisted writing in Swedish parliamentary motions. During the 2022–2023 parliamentary year, the researchers classified 0.1% of motions as AI-assisted overall, while 1.7% contained at least one AI-assisted paragraph. Those figures increased to 1.9% and 4.1% respectively in 2023–2024, followed by 4.7% and 8% in 2024–2025. For 2025–2026, the corresponding figures reached 6.5% and 9.4%.
The study found no statistically significant differences in detected AI use between Swedish and British parliamentarians or between political parties.
The research did not assess whether the content of the motions was factually correct or examine the legal implications of using generative AI. Its focus was on linguistic patterns associated with AI-assisted writing.
Suvanto said the absence of disclosure could raise questions about transparency, particularly because AI-generated material can appear polished even when it contains errors.
“Regardless of the extent to which AI has been used, there may be risks associated with a lack of transparency,” she said. “Texts produced with the help of AI are often very well-written and persuasive, which increases the risk that errors will not even be spotted by the writers themselves.”
The researchers also pointed to broader concerns about the use of artificial intelligence in political communication. A report from the SOM Institute covering Swedish attitudes towards AI found that 88% of respondents were quite or very concerned that false information generated by AI could influence democratic elections.
The detection systems developed for the study are not commercially available and were created specifically to analyse parliamentary texts in Swedish and English. The Chalmers research group has also developed related systems for distinguishing between human-written and AI-generated literary texts.
