The AI Election: How Full Fact is Leveraging New Technology for UK General Election Fact Checking

This year is being dubbed "the year of the AI election," with much focus on the potential negative impacts of new tech like deepfakes and targeted political ads. While Full Fact shares some concerns about these issues, we are also pioneers in using AI positively to tackle misinformation. Here’s how AI is helping us fact check the 2024 UK general election.

Over the past decade, we've invested in technology to make fact checking faster and more effective. Our tools, tested globally in elections in South Africa, Nigeria, Liberia, the Democratic Republic of the Congo and Senegal, are now monitoring the 2024 UK general election. 

They track a wide variety of platforms, including online news, social media, radio, TV and YouTube. This allows our fact checkers to monitor claims by politicians and political parties wherever and whenever they are made, at all times of the day and night. 

To identify the most important claims to fact check, we've developed a claim-type classifier using Google Research's BERT Large Language Model, fine-tuned with our data.  BERT is trained on millions of sentences in over 100 languages, making it highly effective in understanding and processing language. This tool reduces the number of claims fact checkers need to review, filtering them by topic and allowing searches by keyword and speaker.

The tool also checks if new claims match any we've previously fact checked, using another large language model combined with entity recognition. This helps fact checkers respond quickly to repeated false claims, since the fact check has already been completed. 

Full Fact’s team has used our tools to monitor and fact check live events, like party leader debates. The live transcription highlights new claims for fact checking and flags incorrect statements. This allows fact checkers to respond quickly to inaccuracies and insert correct information into the public debate when it matters most.

Fact Checking Party Manifestos

Building on success from the 2019 general election, AI has sped up our fact checking of party manifestos. This week, as each party launched their manifesto, we first used our AI tools to transcribe leaders’ speeches and interviews. Using speech-to-text technology our tools created a convenient text transcript of the event and identified what types of claims were being made, and whether we had previously fact checked them.

Once each manifesto was published, we analysed it with our AI tools. Processing a 100-page document usually takes only a few minutes. This technology enables our team to work at a scale and speed that would be impossible manually, saving time and allowing us to act quickly. 

The analysis also produces lists of the most commonly used words and identifies key phrases central to the document's meaning. We also count the frequency of certain pre-specified words of interest, such as "honesty," "truth," and "trust" as part of our election pledge, and words like "technology" and "AI" to see what proposals are being made in those areas.

For example, this analysis showed that the Liberal Democrat manifesto made 70 claims in the form of predictions, while the Labour Party's claims in this style totaled 576. Here’s a snapshot of what we found this week:

We were also able to extract some key phrases in each of the main parties’ manifestos, using a standard keyphrase extraction algorithm. This combines a measure of how often different phrases appear with a measure of their semantic relatedness and how closely the words appear to each other in the document. The result is a set of key phrases in each manifesto:

• Liberal Democrats: mental health service, mental health professional, criminal justice system, early years education
• Conservatives: new State Pension, free trade agreement, National Living Wage
• Greens: free personal care, nuclear power station, hospital waiting list
• Labour: National Wealth Fund, Great British Energy, clean energy superpower, free breakfast club
• Plaid Cymru: conservative UK Government, Labour Welsh Government, criminal justice system

We also conducted a second stage of analysis using Google’s Gemini Large Language Model, a competitor to ChatGPT, to find key phrases, pledges, and claims in each manifesto. This process, which takes only a few minutes, produces a spreadsheet listing all the key phrases it identifies, along with up to 50 pledges and 50 claims.

Responsible AI Use in Elections and Beyond

Full Fact tools show how AI can be used for good, scaling our efforts to monitor the UK’s General Election and identify false and harmful claims. By providing reliable information quickly during the election, we’ll support voters in navigating the multitude of claims and promises from political parties during this campaign.

While AI has the potential to be a force for good, it can also be misused. We urge all parties to commit to transparent and responsible AI use during elections. We also call on the next government to strengthen society’s resilience to AI-generated misinformation by introducing stronger online safety legislation and increasing resources for media literacy. 

With only two mentions of "misinformation" in all the manifestos, we remain cautious. The real challenge begins after July 4, when we will present our evidence from our election fact checking to the new government, advocating for the necessary changes to ensure the safe and beneficial use of new technology in society.