The UK's New Election Deepfake Detector Works. The Law Behind It Doesn't.

Ahead of the UK's May 2026 local elections, the Electoral Commission ran a quiet experiment: an AI model watching roughly three million public social media posts for signs of election-disrupting deepfakes. It flagged 147. Human reviewers judged ten serious enough to refer to platforms or police. Speaking at the Conservative Party's annual conference this month, chief executive Vijay Rangarajan held that low hit rate up as proof the approach works, and as the case for pushing it further: the Commission is now building toward alerting candidates and parties to a significant deepfake within three hours of it appearing online, which Rangarajan says is roughly how long a post takes to go properly viral. The system is deliberately narrow. It doesn't try to identify who made a fake, since tracing a creator can take weeks or months longer than any election allows; instead it screens posts in near-real time and leaves attribution for afterward, an approach Rangarajan calls "perpetrator neutral." It also only targets content that could disrupt how an election is actually run, like false claims about voting dates or a candidate withdrawing, not the ordinary noise of campaign spin, which the Commission has no mandate to police.

A law written for leaflets

The gap the AI is meant to patch over is legal, not technical, and it predates the internet. Section 106 of the Representation of the People Act 1983, the statute still used to prosecute false statements about candidates, only applies during a campaign's formally "regulated period." In written evidence to Parliament's Foreign Affairs Committee earlier this year, the Commission disclosed that police could not prosecute over a fabricated video of Conservative MP George Freeman precisely because it surfaced outside that window. A detector built to catch a fake within three hours is still operating under a law that, by the Commission's own account, often can't touch what it finds.

If that happens to the British election, that will move a significant amount of votes.

Rangarajan wasn't speaking hypothetically. During Ireland's 2025 presidential race, a deepfake video circulated showing candidate Catherine Connolly announcing her withdrawal and urging supporters to stay home. She won anyway. What turns the episode from cautionary anecdote into a sharper contrast is what happened next: as President, Connolly spent July 2026 signing Ireland's own Regulation of Artificial Intelligence Act into law, making Ireland one of the first EU states to stand up a dedicated national AI authority under the bloc's AI Act. The country that absorbed one of the starkest documented AI-election attacks now has more statutory AI infrastructure in place than the watchdog racing to build a faster alarm for the next one.

The alert can't see everything

The pilot's own blind spots cut the same way. Full Fact, the fact-checking charity that monitored more than 1,000 candidates during the same May elections, found AI-generated campaign videos from a candidate in Glasgow that never triggered a Commission referral, not because the system missed them, but because each clip carried a small "illustrative AI scene" label in its caption, enough to place it outside the pilot's scope. A labeling convention with no legal force decided what counted as disclosed. Rangarajan's more striking admission concerns something no three-hour alert can address at all: what he called "algorithmic bias," signs that at least one social platform's recommendation system skewed political content during a UK campaign. He wouldn't name the platform, and said current law gives the Commission no tool to investigate or remedy it, since nothing in UK statute requires platforms to disclose how their ranking systems treat election content. His proposed fix isn't more authority for his own agency; he argued a separate, better-resourced "sovereign capability" should take on large-scale platform monitoring, an implicit admission that a body built to certify ballots and register donations isn't equipped for it. The clock is already running: the next UK general election must happen by August 2029, against a legal framework from 1983 and 2000 that Rangarajan calls "very out of date," the same conclusion Parliament's own Joint Committee on Human Rights reached in a September report, which found no general transparency duty exists anywhere in UK law for AI systems and gave ministers until mid-November to respond.