Fact-checkers wrestle with how to minimize AI’s problems
For Alan Duke, co-founder of the fact-checking outlet Lead Stories, the challenges of trusting AI in an era of malleable truth were crystallized recently in his hotel room.
Duke had just loaded Llama, an AI chatbot, onto his Mac Mini, and he decided to test it by asking about his longtime Lead Stories collaborator, Maarten Schenk.
The chatbot started by telling him that Schenk was a “fact-checker” — so far so good. Then, Llama’s went off in an unexpected — and inaccurate — direction, Duke told attendees June 19 at the annual GlobalFact conference for fact-checking and counter-misinformation.
Llama told Duke that Schenk “worked” — past tense — for Lead Stories, “a media company that was criticized for promoting misinformation and disinformation on social media platforms.” The outlet, it said, “often features sensationalized or misleading headlines about various topics.” Llama went on to allege that Schenk left Lead Stories in 2020 “due to concerns over the company’s fact-checking practices of the spread of misinformation.”
None of this was true, Duke said.
One of the themes during GlobalFact’s speeches and panels was what to do with AI. It can be a savior for workflow. It can be an obstacle to uncovering the truth. But one way or another, it is affecting much of the universe fact-checkers inhabit.
“In the last 18 months, the revolution has been extraordinary, and the newsrooms across the world have more questions than answers on how to use AI efficiently in their work, and how to protect themselves from the harms that the AI is bringing,” Rawan Damen, director general, Arab Reporters for Investigative Journalism, told attendees.
Rawan Damen, director general of Arab Reporters for Investigative Journalism, at the GlobalFact conference on June 19, 2026. (Louis Jacobson/Poynter)
GlobalFact panels in Vilnius included, “Response, Retrieval & Resistance: Auditing How AI Models Handle Politically Contested Information,” “The Scam You Won’t See Coming: How AI is Rewriting the Fraud Playbook,” “The AI Threat: The ‘Answer Economy’ and the Vector for Chaos,” and in a more optimistic vein, “AI as an Ally: The New Frontier of Fact-Checking Innovation.”
“I’m glad fact-checkers are grappling with both sides of the AI equation,” said Alex Mahadevan, director of MediaWise and Poynter’s AI Innovation Lab. “I see a huge opportunity in it helping us reach more people, debunk more falsehoods and build really cool products — but none of that works without tools we can trust.”
Adrianus Warmenhoven, cybersecurity adviser for NordVPN, told attendees that chatbots build trust with their users by answering non-controversial questions like “how do I open a jar of pickles without using tools.”
“They’ll answer these correctly, because there’s no hallucination involved, no sources involved,” Warmenhoven said. “It’s really simple stuff. … It builds rapport, and it builds trust.”
Then, if you ask the chatbot a more difficult question, “you will trust it, because it was right in all these other answers,” he said.
But this trust may not be warranted, he said. “Those things are really what we have to look at as a society,” Warmenhoven said.
Adrianus Warmenhoven, cybersecurity advisor for NordVPN at the GlobalFact conference on June 19, 2026. (Louis Jacobson/Poynter)
Hurije Mehmeti, an editor with the Hibrid.info of Kosovo, presented findings from research her group did with a series of identical prompts to ChatGPT, a U.S.-based AI tool; DeepSeek, from China; and Alice, from Russia.
The answers, they found, were not neutral. “The responses reflect the ideological structures and data sources upon which they were developed and trained,” the study concluded. “ChatGPT and DeepSeek Chat offer the highest and most consistent accuracy, although they occasionally produce inaccuracies. Alice shows a visible influence, with more refusals to answer, ideological deviations, and use of the Russian language, especially in questions about sensitive topics such as Crimea and Srebrenica.”
Even the user’s location can directly influence the tone, language and narrative of the models, the study found.
On the June 19 panel, Duke asked Warmenhoven how long it would take AI-made fakes to be undetectable — a year? Two years?
Warmenhoven responded, “To be honest, I already think we are at that point.”
Warmenhoven said the ultimate destination for AI mimicry doesn’t have to be perfection; it just has to be “good enough.”
“If you want to spread misinformation or scam somebody, you need to go for ‘good enough,’ ” he said. “You don’t need to have everything perfect.”
About The Author
