Columbia researchers find ChatGPT search frequently misattributes news sources
Researchers at Columbia University's Tow Center for Digital Journalism tested ChatGPT's search feature with 200 quotes from 20 publications and found it misattributed sources in more than 75% of cases, often fabricating source details with high confidence even when blocked from accessing the original articles.
What happened
Columbia University's Tow Center for Digital Journalism tested ChatGPT's search feature by asking it to identify the sources of 200 quotes from 20 publications. ChatGPT returned partially or entirely incorrect responses on 153 occasions, while only acknowledging an inability to respond accurately seven times.
In cases where ChatGPT could not access the original source due to robots.txt restrictions, it frequently fabricated source material instead of informing users that it was blocked. The chatbot misattributed a letter-to-the-editor from the Orlando Sentinel to a Time story, and when asked about a New York Times article on endangered whales, returned a link to a website that had plagiarized the story.
Reported impact
ChatGPT returned inaccurate responses with high confidence, rarely using qualifying language like 'appears' or 'might.' This occurred even for quotes from publishers with licensing arrangements allowing OpenAI's crawlers to access their content.
The errors affected both publications that blocked OpenAI's crawlers and those that granted access through content licensing deals, with more than a third of ChatGPT's replies containing fabricated source material.
Index analysis
The failure to accurately attribute sources or admit uncertainty when blocked from accessing content suggests the search feature conflates paraphrasing web results with verified citation. The system's confident tone masks frequent inaccuracies that could mislead users about the provenance of information.
Governance controls
OpenAI stated that 'misattribution is hard to address without the data and methodology that the Tow Center withheld,' but did not disclose steps to verify source attribution before returning results.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.