AI Fact-Checkers That Lie: Why ChatGPT and Its Peers Fail at Verifying News
AI fact-checking tools like ChatGPT often hallucinate, producing false citations and inaccurate verdicts despite confident tone.Summary
AI fact-checking tools such as ChatGPT, Claude and Gemini are frequently used to verify news claims, yet research shows they often hallucinate—producing false citations, misattributed quotes and confidently incorrect verdicts. Studies from the Columbia Journalism Review, Stanford HAI AI Index and BBC/EBU reveal error rates ranging from 37% to over 60% on news-related queries, with hallucinations stemming from the models's tendency to fill knowledge gaps with plausible fabrications.
These failures persist because models are trained on data that ends months or years earlier, cannot reliably judge recent events, and even web-augmented systems mishandle live information by synthesizing from outdated or erroneous sources. While multi-model setups and retrieval-augmented approaches reduce some errors, they do not eliminate deeper fabrications, and fact-checkers now treat AI as a hypothesis generator, verifying every claim against primary documents, multiple outlets and official records before accepting any AI-generated verdict.
(Source：Webpronews)