Custom detectors find AI-assisted text in almost one in ten parliamentary motions

The design's strength is a negative control that the calendar handed the researchers for free. Every motion in the training set was filed between 2014 and 2020, before the established chatbots existed, so all 1,400 Swedish and 1,400 British documents are human by construction ^[12]. The machine half of each pair was manufactured: a chatbot summarised an original, then wrote a fresh text from its own summary ^[13]. That is 2,800 real motions and roughly 5,600 texts in total ^[19], enough to fit one detector for Swedish and one for English ^[11].

Then the holdout. Motions from 2021 were written when the current AI support was not available ^[16], so a positive there is a false positive by assumption. The account gives the size of that set as "a large number" and does not say whether the single mistake was a text wrongly flagged or one wrongly cleared ^[16]. At the prevalence being reported, that gap matters: when nine motions in ten are clean, even a low false positive rate contributes a noticeable share of the flagged pile. A rise measured with one fixed detector across four parliamentary years ^[20] is not well explained by detector error, which does not drift upward on its own. The exact height of the line is less certain than its upward direction.

The second limit sits in what the positive class contains. The detector learned to recognise prose written from scratch by a chatbot working off a summary ^[13], which is the heaviest use imaginable, while the reported result counts motions containing AI-assisted text in some form ^[17]. How it scores a human draft that a model tidied, or one generated paragraph inside a human document, is not reported, and tidying is precisely the case Suvanto says the EU rules exempt ^[10].

Zero disclosures across a window that closes this spring ^[2] describes a norm rather than a breach. Sweden had no binding labelling rule until the European Commission's transparency requirements took effect on 2 August 2026 ^[7], and those requirements speak to companies and public sector bodies labelling content ^[8]. Suvanto's own reading is that the regulation is rather vague ^[10].

The stated worry is about review rather than volume: AI-assisted text tends to be well written and persuasive, which Suvanto argues makes errors less likely to be caught even by the person filing them ^[4]. This study cannot test that, because the team kept strictly to the language and did not examine factual content ^[9]. Nor does the public-opinion figure do the work it appears to do. That 88% of Swedes are quite or very concerned about AI-generated false information influencing elections ^[5] is anxiety about disinformation, not an expectation that a member label a motion. I would treat the upward trend as established, and the one-in-ten level as an estimate whose error bar has not been published.