Study: AI Ghostwrote Up to 15% of UK and Swedish Parliamentary Texts, and Nobody Told Voters

Politicians write the laws. Increasingly, they're not the ones writing them.

A study out of Chalmers University of Technology in Sweden and Edinburgh Napier University in the UK analyzed 13,565 Swedish parliamentary motions and 4,209 UK parliamentary statements from 2021 through April 2026. Using custom AI detectors trained to tell human writing from large-language-model output, researchers found that 15% of UK parliamentary statements from 2026 contained at least one AI-generated paragraph. In Sweden, 9.4% of motions submitted during the 2025-2026 parliamentary year showed the same, roughly 300 motions.

None of them said so.

The trend line is steep. UK statements with AI-flagged paragraphs went from 3.5% in 2022 to 15% in 2026. In Sweden, according to data from Chalmers, the paragraph-level rate climbed from 1.7% in 2022-2023 to 4.1%, then 8%, then 9.4% by 2025-2026. Looking at whole documents rather than single paragraphs, 6.5% of Swedish motions and 2.1% of UK texts were classified as fully AI-generated in the latest period, per the Sweden Herald's reporting on the same study.

"What is remarkable is not primarily that more and more politicians seem to be using AI, but that no one says they are using AI," said Mattias Wahde, a Chalmers professor involved in the research, according to the Sweden Herald.

Minerva Suvanto, the study's lead author and a Chalmers PhD student, put it more bluntly to Euronews: "Regardless of the extent to which AI has been used, there may be risks associated with a lack of transparency." She added that AI-polished text "is often very well-written and persuasive," which raises the odds that mistakes slip through unnoticed by the very lawmakers whose names go on the document.

Jan Riise, a Swedish MP from the Green Party (Miljöpartiet) who serves as the party's AI spokesperson, told Euronews he wasn't surprised. He said his party uses AI for research and has internal guidance urging caution, and that he personally prefers writing in his own words. But he drew a line: using AI to draft an entire proposal, he said, "is another story" and politicians should label it when they do.

The pattern isn't just European

New York Focus ran a parallel check closer to home, feeding 54 opinion pieces attributed to New York elected officials, published since May 1 in City & State, the Albany Times Union, and the New York Daily News, through the AI-detection tool Pangram. Just over half, 28 of 54, contained text Pangram flagged as AI-generated.

Nine politicians confirmed to New York Focus that AI was used, including Adrienne Adams, the Democratic nominee for lieutenant governor and former New York City Council speaker, whose spokesperson acknowledged AI helped compose her Daily News op-ed. Five said they weren't sure, four denied it, and nine didn't respond. None of the flagged pieces carried a disclosure, even though New York lawmakers passed a bill this year that, if Governor Kathy Hochul signs it, would require news outlets to flag substantial AI content.

There's a fair counterargument here, and it deserves airing. Politicians have used ghostwriters and staff-drafted speeches for as long as there have been politicians. Riise and several officials contacted by New York Focus made essentially that point: outsourcing writing to a chatbot isn't fundamentally different from outsourcing it to a junior staffer, and Pew Research found about half of U.S. adults now use AI chatbots themselves. If a senator's chief of staff can draft remarks without disclosure, why does a large language model doing similar work require a public label?

The answer researchers give is trust, not novelty. Suvanto pointed to the SOM Institute's "Swedish AI Opinion 2023-2025" report, which found 88% of Swedes are quite or very concerned that AI-generated false information could influence elections. A staffer's draft is still checked and owned by a human accountable to voters. An AI system trained on scraped data carries no such accountability, and its fluency, per Suvanto, makes errors harder to catch.

Rules arriving late

The European Union's AI Act transparency requirements took effect August 2, 2026, obligating companies and public bodies to label AI-generated content under certain conditions. Suvanto told Cision the rules are "rather vague" with exemptions for AI used only in editing, and it remains to be seen whether Swedish politicians change their disclosure habits as a result.

California moved on a related but distinct front. Governor Gavin Newsom's desk now has Senate Bill 1159, from state Senator Christopher Cabaldon (D-West Sacramento), which the Legislature passed after a February Los Angeles Times investigation found a consultant named Matt Klink used a platform called CiviClick to generate 20,000 public comments opposing a gas-appliance rule at the South Coast Air Quality Management District. Regulators there scrapped the rule. CiviClick's CEO, Chazz Clevinger, said the company didn't use AI in that campaign but confirmed it's a tool the company offers elsewhere, and he wouldn't say how the 20,000 commenters were identified. A similar episode surfaced at the Bay Area AQMD involving a platform called Speak4, where the San Francisco Chronicle found ten people who said letters submitted under their names weren't ones they wrote. SB 1159 bars anyone from knowingly using AI to fake a real person's engagement with a government agency and clarifies that agencies don't have to treat bot activity as authentic.

Separately, Capitol Hill has its own AI reckoning underway. A nonprofit called CivAI has been demonstrating for congressional offices how quickly AI can assemble detailed personal dossiers, gun ownership, church attendance, abortion clinic visits, from commercial data brokers, according to Politico, which cited both Republican and Democratic staffers alarmed by the demos. Representative Lori Trahan (D-Mass.) called it "a wake-up call for Congress, which has spent years failing to act," noting the demos landed amid the debate over renewing surveillance authorities that lapsed in June.

None of these threads are hypothetical anymore. The open question is whether any legislature, EU transparency rules included, writes disclosure rules with teeth before AI-assisted lawmaking becomes simply how the job gets done.