Nearly half of Americans now turn to artificial‑intelligence tools for research and idea generation, but a growing body of evidence suggests those tools frequently miss the mark. A March 2025 study by the Tow Center for Digital Journalism found that more than 60% of AI‑powered search responses were inaccurate, while a BBC analysis placed the error rate at roughly 45%.

Industry‑wide benchmarks paint a similar picture. In the RealFactBench test, Claude achieved the highest score at 73% accuracy across all metrics. Gemini 2.5 Pro led a later Google‑run assessment with 55.6% accuracy, and most other models, including OpenAI’s ChatGPT, failed to exceed the 50% threshold. The numbers contrast sharply with claims made by the models themselves; when asked about their own performance, ChatGPT cited 90‑96% accuracy on professional‑style exams, yet the cited source was nonexistent.

Human fact‑checkers continue to rely on meticulous, source‑driven processes. WIRED’s fact‑checking team, for example, cross‑checks each claim line‑by‑line, reaches out to primary sources, and subjects findings to legal and ethical review. "You definitely need a human being," says Mark Frankel, head of public affairs at Full Fact, a UK‑based organization that uses AI to triage claims but still depends on people to verify the details.

Angie Holan, head of the International Fact‑Checking Network, acknowledges AI’s utility when it directs investigators to authoritative sources, but warns that the technology’s hallucinations can mislead. "That way you can understand the strengths and weaknesses of these tools," she notes, emphasizing the need for journalists, librarians, and archivists to stay engaged with AI development.

The limitations extend beyond sheer inaccuracy. Large language models repackage existing knowledge, often reproducing content without attribution. In practice, they can generate plausible‑sounding answers that turn out to be scraped or fabricated, as illustrated by the author’s own experience of receiving a vegan cream‑cheese recipe that never existed in the source material.

Despite the hype, researchers remain skeptical about rapid improvements. A 2025 report from the Association for the Advancement of Artificial Intelligence found that 60% of surveyed experts doubted that the factuality problem would be solved anytime soon. As AI continues to evolve, the consensus among fact‑checkers is clear: automation can assist, but it cannot replace the rigorous verification that human expertise provides.

This article was written with the assistance of AI.
News Factory SEO helps you automate news content for your site.