It was 15 years ago. Social media emerged, citizen journalism proliferated, newsrooms struggled and computer science professor Jun Yang had an idea. Could AI lead to faster fact-checking? Yang had been working with large language models – then an emerging technology – and he cared about investigative reporting. Working with Bill Adair, journalism and public policy professor and creator of PolitiFact, Yang created an AI that could fact-check in real time.
“[Politicians] will basically spin the data in a particular way and make an argument,” says Yang. “We’re seeing a lot of cherry-picking and we’re hoping to basically expose that.”
Human fact-checkers have intuition and experience, but an AI has only data. No problem for Yang. His example: If a politician claimed unemployment fell by X percent while they were in office, the algorithm would check their statement against public records to see if it’s accurate. Working with seasoned journalists, Yang converted journalistic thinking to computational procedures with the aim of making their jobs easier.
But then came the 2016 election campaign, and the game changed. Yang and his computational fact-checking changed, too.
“People are not lying with these subtle lies anymore,” he says. “It’s not about numbers anymore … that really changed the way I approached fact-checking.”
He recalls monitoring presidential debates in Adair’s office. When a candidate made a claim, his fact-checker popped up. Yet fact-checks on their own simply didn’t resonate, Yang and Adair have both said in interviews. Maybe a few people in the middle paid attention, but that’s it. From there, Yang pivoted to a type of fact-checking based in calling out lies.
Read More on AI at Duke
“It becomes a very simple natural language processing problem,” Yang says. “Quote-unquote simple.”
Since COVID, Yang has targeted yet another high-stakes public issue: vaccine hesitancy. Reliable, accurate information is not the issue. These days he’s working on educating people who believe misinformation about vaccine safety – a much trickier challenge, Yang notes.
“On the messaging part, I have to say, LLMs have gotten a lot better,” Yang says. “It can see connections that it would take humans a lot of experience to recognize.”
An AI, he says, can make short, effective pro-vaccine arguments based on a person’s affinities and beliefs, but the algorithm isn’t making decisions or talking directly to people. It’s a tool for humans, with their judgment and intuition, to choose to use or not.