The Problem: Scale That Outpaces Human Capacity
Full Fact’s editorial team consists of eight journalists. Against them: an internet’s worth of claims, candidates, and campaign content spread across Facebook, TikTok, X, YouTube, and Instagram — across multiple simultaneous elections in Scotland, Wales, and England.
The core tension is simple. Misinformation spreads in volume. Fact-checking, done properly, requires judgment. No human team of eight can monitor thousands of candidate accounts in real time, scan tens of thousands of images for manipulation, and still produce rigorous, publishable fact-checks.
Something had to give — or something had to scale.
The Tool: Full Fact AI
Full Fact has been applying machine learning to fact-checking since 2016. What began as an internal efficiency project has evolved into Full Fact AI, a modular suite of monitoring and analysis tools now used by over 40 fact-checking organizations working across 30 countries in three languages.
The system operates across multiple media formats and source types.
What It Does
On a typical weekday, Full Fact AI processes approximately 330,000 sentences drawn from online news sites, social media platforms, and video content. The pipeline handles three core functions:
- Claim detection — identifying new, checkable claims from monitored sources
- Claim matching — surfacing repeats of claims that have already been fact-checked
- SynthID scanning — detecting Google’s invisible digital watermark in images and videos to flag potentially AI-generated content
For video content, the system generates transcripts automatically, feeding spoken claims into the same detection pipeline as written text. This matters because campaign content increasingly lives in video format, where manual monitoring is prohibitively slow.
The Setup: Monitoring 1,000+ Candidate Accounts
Going into the Scottish and Welsh parliamentary elections in May 2026, Full Fact significantly expanded its monitoring scope. Using candidate account data from Democracy Club, a digital democracy organization, the team began tracking more than 1,000 social media accounts linked to election candidates across five platforms.
The workflow was deliberately designed to minimize friction.
Claim matches were not buried in a dashboard requiring active retrieval. Instead, they were posted directly into an internal Slack channel, surfacing relevant content in the environment journalists already worked in. This small architectural decision — routing alerts into existing communication infrastructure — reduced the cognitive overhead of acting on the data.
Human editors retained full control over what warranted investigation. The AI surfaced; the journalists decided.
Claim Detection in Practice
The monitoring pipeline flagged an incorrect claim about youth unemployment made by a candidate in Wales — a post the editorial team acknowledged they would likely have missed through conventional monitoring. The claim was fact-checked and published.
This is the practical value of scale: not replacing editorial judgment, but expanding the surface area that judgment can be applied to.
SynthID Detection: 136 Flagged Across 16,514 Assets
Over the course of the May elections, Full Fact scanned 16,514 images and videos attached to candidates’ social media posts. Of these, 136 appeared to carry SynthID watermarks, indicating potential AI generation or editing via Google’s tools.
The majority were unambiguous and uncontroversial — AI-rendered visualizations of proposed construction projects, stylized infographics, illustrative campaign imagery. But a subset warranted closer scrutiny.
The most significant case involved an independent candidate standing in Glasgow. A campaign video showing the candidate meeting voters, visiting schools, and touring a hospital turned out to be entirely AI-generated. The candidate labeled the footage as “illustrative” — representing aspirations rather than actual events. Full Fact identified the video through the SynthID scan. Without the automated detection, the team states they would likely never have seen it.
Post-Election Analysis: 33,000 Posts, Processed Rapidly
After the elections concluded, Full Fact used generative AI tools to analyze more than 33,000 posts from Scottish and Welsh parliamentary candidates. The analysis produced a structured overview of campaign themes: the economy dominated across both nations, while Scottish independence registered as a significantly larger topic in Scotland than in Wales.
This kind of retrospective analysis — producing a structured, comparative snapshot of campaign discourse — would have required weeks of manual work. The AI-assisted approach compressed that into a fraction of the time.
Limitations: What the System Cannot Do
Full Fact is transparent about the boundaries of its tooling.
SynthID detection is only as broad as Google’s watermarking adoption. Content generated by tools that do not embed SynthID — a substantial and growing category — falls outside this detection method entirely. The 136 flagged assets represent what was detectable, not what was AI-generated in total.
Claim matching depends on the quality and coverage of the existing fact-check database. Novel misinformation, or claims that have not previously been checked, will not trigger a match. The system identifies repetition; it does not independently assess truth.
And the entire pipeline requires human editorial judgment at the output stage. The AI reduces the search space. It does not replace the verification process.
The Broader Principle: Small Teams, Larger Coverage
Full Fact’s approach offers a transferable model, particularly relevant for newsrooms preparing for high-stakes election coverage with limited staff.
The architecture rests on three principles that held throughout the May elections:
Reduce friction relentlessly. Routing alerts into Slack rather than a separate dashboard is a small decision with significant behavioral consequences. Tools that require active retrieval get checked less often.
Keep humans in the loop at the decision point. AI surfaces candidates for investigation. Journalists determine what is worth publishing. This division of labor is not a compromise — it is the design.
Use AI for volume; use journalists for judgment. Processing 330,000 sentences a day is not a task that scales with headcount. Deciding which of those sentences represents a story worth telling is not a task that scales with compute.
Takeaway
Full Fact AI is not a fact-checking machine. It is a monitoring and triage system that allows a small, skilled editorial team to operate at a scale that would otherwise be structurally impossible.
The Glasgow campaign video case illustrates the stakes clearly. A piece of AI-generated political content, labeled misleadingly, circulating on Facebook — found not by a journalist scrolling through feeds, but by an automated watermark scan across 16,514 assets. That is what appropriate AI integration looks like in practice: not replacing the reporter, but ensuring the reporter has a fighting chance of seeing what is actually out there.
As AI-generated content becomes cheaper, faster, and harder to detect, the organizations best positioned to hold it accountable will be those that have already built the infrastructure to find it.
Comments (0) No comments yet
Want to join this discussion? Login or Register.
No comments yet. Be the first to share your thoughts!