Detecting undisclosed LLM-generated content in parliamentary texts

Undisclosed use of large language models to draft parliamentary texts in the United Kingdom and Sweden has risen steadily since 2022, according to a new study that trained an interpretable classifier to detect machine-generated prose in official motions [1, 2]. The research, submitted for publication in June 2026, applied a glass-box text classifier to recent parliamentary records after training it on pre-LLM documents and their LLM-generated counterparts [1, 2]. The authors report a consistent increase in undisclosed AI-authored content in both parliaments beginning in 2022 [1, 2]. Unlike journalism or academic writing, where disclosure requirements are often explicit, parliamentary guidelines on AI use remain vague, the paper notes [1, 2]. “In order to maintain transparency and retain public trust, it is generally recommended that parliamentarians should state whether or not they have used AI when writing texts, such as parliamentary motions,” the researchers write [2]. The findings arrive as detection science grapples with the limits of binary human-or-machine classification. A separate study presented at ACL 2025 demonstrated that large language models can iteratively refine legislative amendments to conceal self-serving intent, with identification rates for hidden benefactors dropping by up to 40 percentage points across two optimization trials [3]. That work simulated a lobbyist agent proposing amendments with an overtly altruistic agenda while embedding benefits for a specific company, grounding experiments in real U.S. congressional bills paired with SEC 10-K business descriptions [3]. Other researchers have argued that binary detection is insufficient for real-world scenarios where humans and machines collaborate on text. A 2024 paper proposed a fine-grained classification schema distinguishing fully machine-generated, machine-polished, and human-written text, noting that machine-polishing may be acceptable in some contexts but not in others, such as education [4]. A contemporary benchmark, LLMDetect, introduced tasks for recognizing the specific role an LLM played in content creation and measuring the degree of LLM involvement, finding that fine-tuned pre-trained language models consistently outperformed zero-shot LLM detectors [5]. The Swedish parliament’s vulnerability to undisclosed AI content sits against a backdrop of documented information-manipulation concerns. In 2015, the Swedish Security Service concluded that Russia was using fake news to inflame “splits in society” through propaganda, and Sweden’s Ministry of Defence tasked its Civil Contingencies Agency with countering the threat [6]. The European Parliament’s Committee on Foreign Affairs later warned that pseudo-news agencies and internet trolls were being used as disinformation tools to weaken confidence in democratic values [6].