Image: TechcityAuthority, data from EBU and BBC

A link at the end of a chatbot answer does a lot of work. It makes the answer look checked, as if someone went to the source and confirmed it. ChatGPT, Gemini, Perplexity, Copilot and Grok all attach links like this when they search the web. Several large studies have now tested whether those links actually support what the chatbot says. The results suggest that a citation is often closer to decoration than proof.

What the news studies found

The Tow Center for Digital Journalism at Columbia tested eight AI search tools and published the results in the Columbia Journalism Review in March 2025. Researchers gave each tool an excerpt from a real news article and asked it to name the headline, the publisher, the date and the URL. They ran 1,600 queries across 20 publishers.

Collectively, the tools gave incorrect answers to more than 60% of the queries. Perplexity had the lowest error rate at 37%. Grok 3 got 94% of its answers wrong. More than half of the responses from Gemini and Grok 3 cited fabricated or broken URLs, and 154 of Grok 3's 200 citations led to error pages.

The tools were also rarely honest about being unsure. ChatGPT misidentified 134 articles, according to the study, but signaled a lack of confidence only 15 times across its 200 responses and never declined to answer. The researchers found that the premium versions of some tools gave more confidently incorrect answers than the free ones.

The BBC ran its own test in December 2024 and published it in February 2025. BBC journalists reviewed answers from ChatGPT, Copilot, Gemini and Perplexity to 100 questions about the news, with the assistants asked to use BBC sources where possible. The BBC said 51% of all answers had significant issues of some form, and 91% had at least some issues. It found that 19% of answers citing BBC content introduced factual errors such as wrong numbers and dates. It also found that 13% of quotes sourced to BBC articles were either changed or not in the cited article at all.

The largest study so far came from the European Broadcasting Union and the BBC. Their News Integrity in AI Assistants report brought together 22 public service media organizations across 18 countries and 14 languages. Journalists evaluated 2,709 responses from the free consumer versions of ChatGPT, Copilot, Gemini and Perplexity, generated in late May and early June 2025.

According to the full report, 45% of responses had at least one significant issue, and 81% had some kind of problem when smaller issues were counted. Sourcing was the biggest cause, with 31% of all responses showing significant sourcing problems. Gemini stood out, with significant sourcing issues in 72% of its responses. The report put ChatGPT at 24% and Perplexity and Copilot at 15% each. It also said 42% of Gemini responses provided no direct source links at all.

There was some improvement. On the BBC's own questions, the share of answers with significant issues fell from 51% in the first round to 37%. That is better, but it still means more than a third of answers had serious problems.

Fake citations outside the newsroom

The problem is not limited to news. In law, the best known early case is Mata v. Avianca. A federal judge in New York sanctioned two lawyers in June 2023 after they filed a brief with six fictitious case citations generated by ChatGPT. The lawyers and their firm were ordered to pay a $5,000 fine, Reuters reported.

Cases like that are no longer rare. Legal researcher Damien Charlotin keeps a public database of court decisions where a party relied on hallucinated content, most often fake citations. At the time of this writing, the database listed 2,217 cases identified so far. Charlotin notes that it is a work in progress and covers decisions where courts found or implied that AI generated material was used.

Academic research has the same issue. AI detection company GPTZero said in January 2026 that it scanned 4,841 papers accepted at NeurIPS 2025, one of the biggest AI research conferences. It said it found 100 hallucinated citations that it confirmed as fake, spread across more than 50 papers. GPTZero sells a citation checking tool, so the finding comes from a company with a commercial interest. NeurIPS told Fortune that even if 1.1% of papers had incorrect references due to AI, the content of those papers was not necessarily invalidated.

Why the links go wrong

It helps to know that a chatbot answer with links usually comes from two separate steps. In the first, the system searches the web and pulls in some pages. In the second, the language model writes an answer. The model is predicting likely words, not copying facts from a page, so the text can drift away from what the source says.

In some setups, links are matched to the answer after the text is written. That can produce a citation that is on the right topic but does not contain the specific claim it sits next to. The EBU report counted cases where a claim had a source attached that did not support it as a sourcing failure.

Access is another problem. Many publishers block AI crawlers or keep stories behind a paywall. The Tow Center found that tools still answered questions about blocked content and often pointed to copies of the story on other sites. Perplexity Pro cited syndicated versions of Texas Tribune stories in three of ten cases, and ChatGPT cited a Yahoo News copy of a USA Today article even though USA Today blocks its crawler. Licensing deals did not fix this. ChatGPT correctly identified only one of ten San Francisco Chronicle excerpts, even though the Chronicle's owner has a deal with OpenAI.

Some links are simply invented. A model can produce a web address that looks right, with a real domain and a plausible path, that leads nowhere. The Tow Center also found that Grok 2 often linked to a publisher's homepage instead of the actual article.

Finally, a chatbot sounds equally sure whether it is right or wrong. The BBC noted in its report that audiences are more likely to trust an answer when an AI assistant cites a trusted brand like the BBC, even if the answer is wrong. A confident tone and a familiar logo are not evidence.

A check that takes under a minute

Nobody needs to give up on AI search to use it more safely. A quick routine catches most of the problems these studies describe.

Open the link. If it leads to an error page or a homepage, the citation does not support the claim.

Find the exact sentence. Use your browser's find feature to search for a key number, name or phrase from the answer. If you cannot find it on the page, treat the claim as unverified.

Check the date and the publisher. An old story can be cited as if it were new, and the site in the link may not be the outlet the chatbot named.

Watch for copies. A story on an aggregator may be a syndicated version of the original. Go to the original publisher when it matters.

Be wary of quotes with no link. The BBC found altered quotes and quotes that did not exist in the cited article. If a chatbot quotes a person, find that quote in a real source before repeating it.

What the AI companies say

When the Tow Center asked for comment, only OpenAI and Microsoft responded. OpenAI said it supports publishers by helping 400 million weekly ChatGPT users discover quality content through summaries, quotes, clear links and attribution. "We've collaborated with partners to improve in-line citation accuracy and respect publisher preferences, including enabling how they appear in search by managing OAI-SearchBot in their robots.txt. We'll keep enhancing search results," the company said, as published by CJR.

Microsoft told CJR that it respects the robots.txt standard and honors the directions of websites that do not want their content used with its generative AI models. Neither statement addressed the specific error rates the researchers found.

The takeaway

Citations in chatbot answers are useful. They give you a place to start checking. But across the Tow Center, BBC and EBU studies, a large share of answers either misstated the source, pointed to the wrong place or had no usable source at all. Courts and academic reviewers are running into the same problem with invented references.

The safest way to read an AI answer is as a lead, not a finding. If a claim matters, click through, find the sentence and confirm it says what the chatbot says it does.