A Contradiction in Practice
A recent exploratory survey has uncovered a fascinating trend among Europe’s professional fact-checkers: they are significant users of generative AI (GenAI) tools like ChatGPT, even while expressing very low trust in the outputs these systems provide.
A study involving 47 fact-checkers from 24 different European countries revealed this peculiar relationship. On one hand, the technology is being integrated into daily workflows. On the other, the core principles of fact-checking—accuracy, transparency, and contextual reasoning—are seen as fundamentally at odds with how these AI models operate. This isn’t just a European phenomenon; fact-checkers globally are grappling with the dual nature of AI as both a potential solution and a part of the problem in the fight against misinformation. The central question this raises is not whether they use AI, but how and why they use a tool they inherently distrust.
The Lure of Efficiency
The primary driver behind the adoption of generative AI in newsrooms and fact-checking organizations is not a quest for truth, but a search for efficiency. Fact-checkers are not asking ChatGPT whether a politician’s claim is true. Instead, they are using it for a range of secondary, time-consuming tasks. This includes summarizing lengthy reports, rephrasing complex information for different audiences, brainstorming headlines for their debunks, or even drafting social media posts and scripts for videos. Henrik Brattli Vold, a senior adviser for the Norwegian fact-checking organization Faktisk Verifiserbar, noted that the time from an idea to a prototype is significantly shorter when using these tools. In a world where false information spreads at lightning speed, the ability to accelerate the non-verification parts of the job is a massive advantage, allowing human experts to focus their limited time on the critical work of actual verification.
A Deep and Justified Trust Deficit
The lack of trust is not just a vague feeling; it’s rooted in the fundamental architecture of large language models. Experts point out that ChatGPT’s crucial flaw is that it doesn’t know when it doesn’t know something, leading it to invent information—a phenomenon known as 'hallucination'. For a fact-checker, this is the ultimate deal-breaker. The models are trained on vast swatges of the internet, which is rife with the very misinformation they are tasked with fighting. This means the AI can inadvertently generate content that is inaccurate or misleading, but present it in a convincing, authoritative tone. Furthermore, the lack of transparency—the 'black box' nature of how AI reaches a conclusion—is antithetical to the fact-checking process, which demands a clear, verifiable trail of evidence.
A Human in the Loop
The consensus among fact-checking professionals is clear: AI is a tool to augment, not replace, human expertise. The model is one of a 'human-in-the-loop,' where the technology acts as an assistant that can perform specific, delegated tasks but is never the final arbiter of truth. Researchers have found that while the public may trust AI for simple, large-scale scanning tasks, they still rely on human judgment for more nuanced fact-checking that requires piecing together evidence and understanding context. Fact-checkers use AI tools for ancillary tasks, but the core work—corroborating evidence, evaluating sources, and making a final judgment on a claim’s veracity—remains a deeply human process that requires critical thinking and contextual awareness that AI currently lacks. The goal is to create a collaborative relationship where AI handles the grunt work, freeing up journalists to do the high-level analysis.
