ChatGPT Search is meant to help users find and understand information on the web. But a study from Columbia University's Tow Center for Digital Journalism raises a direct concern for newsrooms: the system can point readers to the wrong source, misstate what a source says, or present a fabricated citation with confidence.
The study tested news citations at scale
The researchers tested 200 news citations from 20 different publishers. In 153 cases, ChatGPT gave source information that was wrong or partially incorrect.
That result matters because citations are not a decorative feature in news search. They are how readers check whether a claim is grounded in reporting, whether the original publisher is credited, and whether the AI system is actually pointing to the material it appears to summarize.
The study did not find that the problem was limited to one category of publisher. According to the source, the errors appeared for both OpenAI partner publishers and publishers that block ChatGPT from accessing their content.
Access agreements did not prevent errors
One of the most important findings is that direct relationships with OpenAI did not guarantee accurate treatment. The source article says that even publishers with OpenAI partnerships and approved content access were affected.
The New York Post and The Atlantic are named as examples of publishers whose content was often misquoted or misrepresented by the AI system. That detail changes the stakes of the issue. If access arrangements do not reliably protect accuracy, publishers cannot assume that licensing or partnership alone solves the citation problem.
For publishers, accurate attribution is part of the product. A reader who sees the wrong source may not know where reporting originated. A reader who sees a misquote may come away with a distorted view of the article. A reader who sees a confident answer may not realize that the citation behind it is unreliable.
ChatGPT rarely said it could not find the source
The researchers also found that ChatGPT admitted it could not find a source in only seven cases. That is a small number compared with the 200 citations tested, and it points to a broader product issue: the system appeared more likely to answer than to clearly state uncertainty.
The researchers described the behavior this way:
"Eager to please, the chatbot would sooner conjure a response out of thin air than admit it could not access an answer,"
The source article says the false sources were presented with complete confidence, without signs of uncertainty. For users, that is especially difficult to evaluate. A hesitant answer can invite checking. A confident answer can look settled, even when it is not.
This is a central problem for AI search. When a search interface gives a link, readers may treat it as evidence. If the link is wrong, copied, or only partly connected to the claim, the interface can create a false sense of verification.
Copied content creates another risk
The study also found cases where ChatGPT pointed to copied content instead of the original publisher. In one example involving a New York Times article, the system linked to a website that had copied the full article without attribution.
That kind of error is different from a simple broken citation. It can shift attention away from the original reporting and toward an unauthorized copy. It also makes it harder for publishers to receive proper credit for their work.
Mat Honan, editor-in-chief of MIT Technology Review, summed up the publisher concern in the source article:
"As a publisher, that’s not something you want to see," said Mat Honan, editor-in-chief of MIT Technology Review. "But there’s so little recourse."
The phrase "so little recourse" captures the practical frustration. The researchers concluded that publishers currently have no way to ensure ChatGPT Search displays their content correctly, whether they work with OpenAI or not.
OpenAI says it is working on citation accuracy
OpenAI responded to the findings by saying ChatGPT directs 250 million users to high-quality content weekly. The company also said it is working with partners to improve citation accuracy.
That response highlights the scale of the system. If ChatGPT is sending large numbers of users to content every week, citation quality becomes more than a technical detail. It affects how readers discover journalism, how publishers are credited, and how trust is assigned across the web.
The complete study results are available on GitHub, according to the source article. For now, the central takeaway is clear: ChatGPT Search can produce inaccurate citations even when the publisher has a relationship with OpenAI, and publishers do not yet have a reliable way to control how their work appears in those results.
For readers, the safest lesson is to treat AI-generated citations as pointers that still need checking. For publishers, the study adds pressure to a larger question: how can news organizations protect attribution and accuracy when AI search systems become part of how people find their work?