Study: AI chatbots usually cite incorrect sources

Popular AI services are not very good at locating the correct original source, according to a new study from Columbia Journalism Review’s Tow Center for Digital Journalism.

In the study, researchers selected 10 articles from 20 different publishers and then manually selected quotes from them to use in their queries. After each chatbot got the quotes, it was asked to identify the corresponding article’s title, original publisher, and publication date.

The researchers deliberately chose quotes that would give the correct original source if typed into the Google search engine.

In total, eight different AI chatbots were tested and, on average, they produced the wrong source 60% of the time. Perplexity performed best — and still got the citation wrong 37% of the time. The worst performer was Grok 3, which was wrong 94% of the time.

The researchers note that while most of the AI tools produced incorrect answers, they still presented them with great confidence. This was particularly true for paid versions of the AI chatbots. The researchers also found evidence that the AI chatbots’ web spiders often ignored publishers’ paywalls they were supposed to respect.