Regional elections in Spain there’s still almost four months to go, but Irene Laraz and her team at Newtral are ready to roll. Every morning, half of Lara’s team at the Madrid-based media company compiles a schedule of political speeches and debates, preparing to fact-check politicians’ claims. The other half, debunking misinformation, scans the web for viral lies and works to infiltrate the groups that spread the lies. Once the May elections are over, national elections are due to be called before the end of the year, which is likely to trigger a flurry of lies online. “It’s going to be pretty tough,” Laraz says. “We are already preparing.”
The spread of misinformation and propaganda online has meant an uphill battle for fact-checkers around the world, who have to sift through and verify vast amounts of information during complex or fast-moving situations such as the Russian invasion of Ukraine, the Covid-19 pandemic, or election campaigns. . This task has become even more difficult with the advent of chatbots that use large language models, such as OpenAI’s ChatGPT, which can generate natural-sounding text with the click of a button, essentially automating the production of disinformation.
Faced with this asymmetry, fact-checking organizations are having to build their own AI-based tools to automate and accelerate their work. It’s far from a complete solution, but fact-checkers hope these new tools will at least prevent the gap between them and their adversaries from widening too quickly as social media companies scale back their own moderation operations.
“The race between fact-checkers and those they fact-check is uneven,” says Tim Gordon, co-founder of Best Practice AI, an AI strategy and management consultancy, and a trustee of the UK’s fact-checking charity.
“Fact checkers are often small organizations compared to those who deal with disinformation,” says Gordon. “And the scale of what generative AI can produce and the pace at which it can do it means that this race is only going to get harder.”
Newtral began development of its multilingual artificial intelligence language model ClaimHunter in 2020, funded by profits from its television division, which produces fact-checking political shows and documentaries for HBO and Netflix.
Using Microsoft’s BERT language model, ClaimHunter developers used 10,000 utterances to train the system to recognize sentences that appear to include statements of fact, such as data, figures or comparisons. “We trained the machine to play a fact-checking role,” says Ruben Miguez, chief technology officer at Newtral.
Simply identifying the claims of political figures and social media accounts that require verification is a difficult task. ClaimHunter automatically detects political claims on Twitter, while another program transcribes politicians’ video and audio reports into text. Both identify and highlight claims that contain claims relevant to public life that can be proven or disproved – such as non-ambiguous claims, questions or opinions – and flag them for neutral fact-checkers to check.
The system isn’t perfect and occasionally marks opinions as fact, but its mistakes help users constantly relearn the algorithm. This reduced the time needed to identify claims worth checking by 70 to 80 percent, Migues says.