Full Fact is battling AI-generated elections content with AI tools of its own

As campaign videos go, it was a classic. To the sound of a soulful pop ballad, a smiling politician meets voters in the street, chats to school children and visits a hospital.

The only problem? The scenes shown in the video weren’t real. They were AI-generated, and shared and labeled as “illustrative” by an independent candidate standing in Glasgow in the Scottish parliamentary elections which took place last month. The video, along with another similar clip also shared on Facebook, “represent my goals — things I aspire to do — rather than past events,” the candidate told us.

Welcome to the world of fact checking in 2026, where increasing use of AI in lots of different contexts is throwing up fresh challenges and questions for journalists and fact checkers alike.

This has been most keenly felt at Full Fact, the U.K.’s independent fact-checking nonprofit, as we’ve covered recent elections in England, Scotland, and Wales. We’ve found that AI imagery is no longer a hypothetical factor: It’s being used, and in increasingly complicated, sometimes surprising ways. But at the same time, we’ve been able to use AI in new ways ourselves to confront the challenge, to scale our monitoring and find new ways to target our work.

Full Fact, an organization of 34 people, includes a dedicated AI team — a group of data scientists and software engineers who work alongside the eight journalists on our editorial team. We’ve been doing this for years, using machine learning to improve and scale our fact-checking since 2016.

In recent years we’ve developed Full Fact AI, a suite of tools that help monitor claims from a wide range of sources — including online news sites, social media and video platforms. The tools can help identify new claims that might be important to verify and find repeats of claims that have already been fact-checked. On a typical weekday, our tools now process about a third of a million sentences in total — and have been used by over 40 fact-checking organizations working in three languages across 30 countries. (You can find out more here.)

At Full Fact, these tools are already fully integrated into our newsroom’s workflow. For instance, each week we use a live transcript of Prime Minister’s Questions to alert us to repeat claims as we fact-check it in real time.

But going into a busy election period, we stepped this up. Based on data collected by the digital democracy organization Democracy Club, we started monitoring more than a thousand Facebook, TikTok, X, YouTube, and Instagram accounts linked to candidates in the Scottish and Welsh parliamentary elections (plus a few mayoral contests in England).

Claims made via these channels — including in videos, for which our AI tools provided transcripts — were then matched against previously published fact-checks. Our journalists were able to search the claims, and we also created a feed posting claim matches directly into an internal Slack channel to minimize friction and maximize our use of the data.

Collating posts in this way helped us identify some fact-checkable posts we might otherwise have missed — for example, we spotted an incorrect claim about youth unemployment from a candidate in Wales.

But crucially, we were also able to scan the posts for evidence of SynthID, the invisible digital watermark that indicates an image may have been created or edited with Google’s AI tools. Over the course of the May elections we scanned 16,514 images or videos attached to candidates’ social media posts, and identified 136 that appeared to have watermarks.

Most of these were obviously and non-controversially AI-generated — such as AI images of yet-to-be-built construction projects, or infographics. But some were worthy of further investigation — such as the Glasgow candidate’s “illustrative” video, which we spotted this way and would otherwise likely never have seen.

While our AI monitoring led directly to us writing some fact-checks, it also gave our small editorial team much greater visibility over what was being talked about online, by surfacing a range of different claims and posts that might otherwise not have been picked up. (And after the election, we were able to use generative AI tools to quickly analyze over 33,000 posts from Scottish and Welsh parliamentary candidates, giving us a unique snapshot of the topics they spoke to voters about in the campaign’s final days — the economy dominated, while independence was a much bigger issue in Scotland than Wales.)

For other newsrooms grappling with similar problems — particularly as U.S. journalists brace for the midterm elections later this year — the principle of integrating AI monitoring into editorial workflows to enable small teams to cover much more (online) ground may be a useful one. Reducing friction in the process wherever possible but still having humans very much in the loop to decide what is worthy of attention was key, we found.