Academics from the UK and Sweden have deployed specialist models which found that more than one in seven documents written by MPs has been aided by the use of LLMs
The use of artificial intelligence tools by MPs has quadrupled in four years – but few admit to using it, according to a new study.
Researchers used specially designed AI detectors to analyse motions presented at the UK Parliament. They concluded that an increasing number of submissions are being written with input from Large Language Models (LLMs) – but that none of the MPs involved had openly disclosed their use of AI.
The claim is based on analysis of 4,200 separate submissions made since 2022. The researchers found 243 contained AI-assisted content, with the percentage increasing year on year from 3.5% at the start of the study period to 15% this year.
No “significant differences” between political parties were found. The study was carried out by researchers from Edinburgh Napier University and the Chalmers University of Technology in Sweden.
A similar look at the Swedish Parliament found at least 10% of text contained at least one AI-generated paragraph in 2025-26.
Andrea McGlinchey, associate researcher at Edinburgh Napier University, said: “While our research does not address regulatory frameworks and focuses only on evidence of undisclosed AI usage in the language patterns of parliamentary texts, our results suggest that greater transparency is required.”
Related content
-
Parliament seeks £140k digital chief to lead three-year tech strategy
-
Parliament buys £340k open-source intel tool to monitor threats The UK Government’s Artificial Intelligence Playbook recommends that officials should be open with the public about where and how algorithms and AI systems are being used in official duties. However, the guidance does not cover members of parliament.
Dr Peter Barclay of Edinburgh Napier University’s computer science group said research indicated that AI “may change the intended tone of a document or even potentially introduce outright errors”.
However, the team said these issues may not be detected by the person using the AI tool.
Researchers trained one detector for English and one for Swedish, starting with around 1,400 parliamentary motions from each country written between 2014 and 2020 – before generative AI tools were available.
To generate equivalent texts using AI, they first had a chatbot summarise each original text and, based on that summary, create a new text of its own. Using large datasets containing labelled texts that were known to be either human-written or AI-generated, the researchers then created highly accurate classifiers able to determine where AI had been used somewhere in the writing process.
The study used an ‘interpretable’ classifier developed at Chalmers University of Technology, meaning that the results can be explained by examining which words and combinations of words were used to identify AI-assisted texts – but this identification depends on overall statistical patterns, not just the presence of one or two telltale phrases.
A version of this story originally appeared on PublicTechnology sister publication Holyrood