Truths in “AI Search Has A Citation Problem”

AI Search Has a Citation Problem—and It Matters More Than You Think

Generative AI tools, like ChatGPT and Perplexity, have surged in popularity, rapidly changing how millions of users access information. Nearly one in four Americans now turn to AI tools over traditional search engines. While convenient, these tools often present significant challenges, especially regarding citation accuracy and reliability, posing major concerns for journalism and public trust.

In a recent study by the Tow Center for Digital Journalism, researchers evaluated eight popular AI-powered search tools. They discovered disturbing trends:

• Frequent inaccuracies: More than 60% of queries resulted in incorrect answers.
• Confident misinformation: AI chatbots consistently presented wrong information without indicating uncertainty, misleading users who trust authoritative responses.
• Fabricated citations: Tools regularly create fake URLs or link to non-existent pages.
• Misattributed content: Chatbots often credited syndicated or copied content rather than the original source, harming publishers by reducing traffic and revenue opportunities.
• Ignored crawler restrictions: Several chatbots seemed to retrieve and use content despite explicit restrictions in publishers' robots.txt files.

Interestingly, premium AI products—often trusted more due to their cost and advanced capabilities—were no better. They answered more questions correctly overall but also confidently provided incorrect answers more frequently than free versions.

The implications are significant. As users increasingly turn to generative AI tools for daily news and information, the potential for misinformation and diminished transparency grows exponentially. Publishers face reduced referral traffic, misattribution, and financial losses, while users risk being misled without realizing it.

Although this study highlights critical flaws, it’s essential to recognize that AI search technology is evolving rapidly. Improvements in accuracy, transparency, and citation are possible and, indeed, expected.

As users, publishers, and technologists navigate this landscape, a balanced, proactive approach will be essential. Understanding both the risks and the potential benefits of AI-powered search can help ensure these powerful tools serve journalism and society responsibly.

Below is Substantiated Truths & Bias, The Rest I'll Leave To You.

Substantiated Truths in Columbia's Article

• Rising use of AI chatbots for search: A significant share of the public is now using AI tools like ChatGPT in lieu of traditional search engines – roughly one in four U.S. adults have tried ChatGPT for information, reflecting the rapid adoption of AI search technologies (McClain, 2024). This growing usage makes the integrity of AI search results a pressing issue for news and information dissemination (AI Search Has A Citation Problem - Columbia Journalism Review) (Americans increasingly using ChatGPT, but few trust its 2024 election information | Pew Research Center).
• High incidence of inaccurate answers: Generative AI search engines frequently provide incorrect or hallucinated answers. In a controlled study, over 60% of queries were answered inaccurately by AI chatbots, underscoring a widespread factual reliability problem in current AI search outputs (Narayanan Venkit et al., 2024) (AI Search Has A Citation Problem - Columbia Journalism Review) (AI search engines often make up citations and answers: Study). Such findings align with broader research on large language models, which often produce plausible-sounding but false information (“hallucinations”) in response to queries (Huang et al., 2023).
• Confident delivery of misinformation: These AI systems tend to present information in a confident, authoritative tone even when the content is incorrect or unsubstantiated, which can mislead users. Studies have noted that LLM-based answer engines seldom express uncertainty or caveats – for example, ChatGPT almost never admitted it couldn’t find an answer in the cited study – making it hard for users to distinguish accurate responses from confident misattributions (Shah & Bender, 2024; Narayanan Venkit et al., 2024) (AI Search Has A Citation Problem - Columbia Journalism Review) (AI search engines often make up citations and answers: Study). This unearned confidence can give users a false sense of security in the AI’s answers (Kadavath et al., 2022).
• Fabrication of sources and citations: “AI Search Has A Citation Problem” highlights that AI chatbots often fabricate sources or citations – a phenomenon well-documented in academic evaluations. For instance, ChatGPT has been shown to invent references that look plausible but don’t actually exist: in one study of medical queries, 47% of ChatGPT’s provided references were found to be entirely fake, and only about 7% were fully correct (Bhattacharyya et al., 2023). Similarly, a psychology study found false citation rates ranging from 6% to 60% when ChatGPT was asked for scholarly references (MacDonald, 2023). Such empirical evidence reinforces the article’s claim that AI tools frequently generate bogus citations instead of reliable sources (AI Search Has A Citation Problem - Columbia Journalism Review) (ChatGPT hallucinates fake but plausible scientific citations at a staggering rate, study finds).
• Misattribution of news content: The article reports that generative search tools often misattribute content – crediting the wrong source or a syndicated copy rather than the original publisher. This is a substantiated concern: the Tow Center study found, for example, that the DeepSeek AI misidentified the source of news excerpts 115 out of 200 times, meaning a majority of its attributions were incorrect (AI Search Has A Citation Problem - Columbia Journalism Review). Independent investigations confirm such misattributions; e.g., Forbes publicly accused the Perplexity AI search engine of summarizing its articles without credit, highlighting that AI tools can repurpose news content without proper attribution (Paul, 2024). Misplacing credit not only misleads readers about the information’s origin but also undermines the original journalists and outlets (Shah & Bender, 2024).
• Failure to link to original sources: Even when an AI chatbot ostensibly finds the right article, it often fails to provide a direct link to the original source, instead pointing users to secondary sites or nothing at all (AI Search Has A Citation Problem - Columbia Journalism Review). All eight AI search tools tested showed a tendency not to drive traffic to publishers’ own websites. This claim is backed by the study’s data and aligns with news industry observations that generative answer engines can result in “zero-click” behavior (users get answers without visiting source sites) – a trend previously seen with search engine featured snippets (Barry et al., 2022). Such behavior means that publishers receive less referral traffic, corroborating the article’s warning that AI search could cut off the audience flow to original news providers (Maher, 2024).
• Frequent broken or fabricated URLs: The Tow Center analysis revealed that many citations given by AI search bots lead to dead ends. For example, more than half of the source links provided by Google’s Gemini and xAI’s Grok 3 were broken or non-existent URLs that returned errors (AI Search Has A Citation Problem - Columbia Journalism Review). In 154 out of 200 test cases, Grok 3 produced a citation link that went nowhere (AI Search Has A Citation Problem - Columbia Journalism Review). This finding is corroborated by other research noting that large language models often output superficially credible web addresses or DOIs that are, in fact, invalid (Bhattacharyya et al., 2023). Such fabricated links hinder users from verifying information and illustrate a concrete aspect of the “citation problem” described in the article.
• Licensed content isn’t cited accurately: The article points out that even formal content licensing deals between AI companies and news publishers have not ensured accurate attribution of those publishers’ material. For instance, OpenAI’s partnership with the San Francisco Chronicle did not stop ChatGPT from misidentifying or failing to properly cite that paper’s articles in Nine out of Ten tests (AI Search Has A Citation Problem - Columbia Journalism Review). This is a truth reflected in the data: having privileged access to content (through licensing or permission) does not automatically translate to correct citation in generative answers. Early evidence suggests that technical challenges in AI retrieval and synthesis still lead to missing or wrong citations, an issue acknowledged by publishers in such partnerships (Howard, as cited in Ja?wińska & Chandrasekar, 2025).
• Expert consensus on transparency issues: The concerns raised in the CJR article echo those of AI and information retrieval experts. Researchers Shah and Bender (2024) note that large language model-based search interfaces “take away transparency and user agency” and can amplify biases, producing answers that sound authoritative but lack reliable grounding. This academic perspective substantiates the article’s broader point that AI search tools pose new challenges to the transparency and accountability that traditional search engines offer (e.g., clear source links) (AI search engines often make up citations and answers: Study) (AI Search Has A Citation Problem - Columbia Journalism Review). The fact that multiple experts and prior studies (e.g., a November 2024 ChatGPT study by the same authors) independently report confident misinformation, faulty attribution, and inconsistent retrieval confirms that the problems identified are real and systemic, not one-off quirks (AI search engines often make up citations and answers: Study).
• Potential harm to news media and public trust: The article correctly highlights risks both to news producers and consumers. When AI systems present news information without proper credit or context, news organizations stand to lose web traffic and revenue – a concern documented by media industry groups (Coffey, 2023, as quoted in Paul, 2024, “we cannot monetize our valuable content… This could seriously harm our industry”). At the same time, the spread of unverified or misattributed information can erode audience trust in both the AI platforms and the news sources wrongly cited (AI Search Has A Citation Problem - Columbia Journalism Review). This dual harm is supported by studies of misinformation, which show that credibility suffers when information is delivered out of context or without clear sources (Mena, 2020). In short, the article’s emphasis on these stakes is well-founded in light of current research and expert opinion on AI-driven information systems.

Inaccuracies or Biases in Columbi'as Articles & Claims

• Overgeneralization from limited samples: The article declares that “all” eight generative search engines tested are bad at citing news, then extrapolates this as an industry-wide problem. This broad conclusion may be overstated, given the study’s sample size and specific test conditions. In research methodology, drawing sweeping generalizations from eight products is risky – results may not hold for other AI systems or future updates (Yin, 2014). Even the authors acknowledge in their limitations that findings “are not intended to be extrapolated to all models” (AI Search Has A Citation Problem - Columbia Journalism Review), yet the article’s title and tone suggest a universal failing. This discrepancy points to a bias of overgeneralization, a common pitfall where case-study results are treated as broadly definitive (Sim et al., 2018).
• Atypical query scenario: The testing method the article describes – providing chatbots with verbatim excerpts from news articles and asking for the source – is not a typical use case for an average user. Most real users do not paste long passages into search engines; instead, search queries tend to be short (often just 2–3 keywords on average) and phrased as questions or keywords (Jansen & Spink, 2006). By using exact excerpts guaranteed to surface the source in Google’s index (AI Search Has A Citation Problem - Columbia Journalism Review), the study sets a very high bar that essentially tests the bots’ ability to perform precise document retrieval. This could exaggerate the failure rate of AI search since these systems are not primarily designed as plagiarism detectors. In other words, the evaluation might be biased against the chatbots by emphasizing a task that is trivial for traditional search but not how generative AI is normally expected to be used, thereby making the AI look worse at “citation” than it would under more typical question-answer scenarios.
• Publisher-centric perspective: The article’s framing is heavily oriented toward the interests of news publishers – focusing on the loss of traffic, proper attribution, and content rights. This perspective, while important, introduces bias by downplaying the user experience aspect. For example, the benefits that AI search might offer to users, such as quick summarized answers or convenience, are largely ignored. Academic discussions of search technology stress balancing stakeholder interests (Nelson, 2021), but the CJR piece predominantly reflects the concerns of journalists and publishers. It describes generative search as “cutting off traffic flow to original sources” (AI Search Has A Citation Problem - Columbia Journalism Review), implicitly valuing clicks to publishers over the possibility that users may still get accurate information without visiting a site. This one-sided viewpoint may stem from the authors’ journalism background, but it risks presenting AI search as purely negative without acknowledging potential improvements in information access for users.
• Implied wrongdoing without definitive proof: The article suggests that multiple chatbots “bypassed” publishers’ Robots Exclusion Protocol directives – essentially accusing them of ignoring websites’ no-crawl instructions (AI Search Has A Citation Problem - Columbia Journalism Review). However, the evidence for deliberate bypass is circumstantial. The authors themselves note there are other ways the models might have obtained the information (e.g., via references on allowed sites or pre-existing training data) (AI Search Has A Citation Problem - Columbia Journalism Review). By leaning into the narrative that these AI companies possibly violated web-crawl norms, the article betrays a bias toward assuming malfeasance. In reality, determining whether a large language model “ignored” robots.txt is complex – it’s possible the content was ingested before the block or provided through licensed feeds. Without direct data on the bots’ crawling behavior, the claim remains speculative. The Reuters report on this issue shows even experts had to investigate carefully; it found Perplexity “likely” bypassed robots.txt, but stopped short of conclusively proving intent (Paul, 2024). Thus, the article’s confident wording (“seemed to bypass”) may mislead readers by asserting impropriety that hasn’t been fully verified.
• Selective use of expert opinion: The piece amplifies voices critical of AI search (citing scholars like Chirag Shah and Emily Bender, who have a well-known skeptical stance on large language models) (AI search engines often make up citations and answers: Study) (AI Search Has A Citation Problem - Columbia Journalism Review), but it does not equally present perspectives from the AI or tech community who might see these tools in a more positive or nuanced light. By only quoting critics who emphasize the drawbacks, the article’s sourcing exhibits confirmation bias – seeking out commentary that reinforces its negative findings. Absent are viewpoints from, say, information science experts who might highlight user benefits or from the AI developers beyond boilerplate responses. This imbalance can skew the reader’s impression, making it seem as if virtually all authorities agree generative search is hopelessly flawed, when in fact, there is an active debate in academia and industry about how to improve these systems (e.g., Xu et al., 2023 discuss methods to increase citation accuracy in LLMs). The lack of counter-arguments or any mention of ongoing advancements tilts the article toward a one-dimensional critique.
• “Premium vs free” model confusion: The article notes an interesting finding that paid “premium” versions of chatbots (like Perplexity Pro or Grok 3) got more answers correct but also made more errors than their free counterparts (AI Search Has A Citation Problem - Columbia Journalism Review). It frames this as a paradoxical failure of premium models. However, this portrayal could be misleading without additional context: the higher error count is partly because those advanced models attempted to answer more queries rather than refusing. In other words, a premium AI might take on a harder question instead of saying, “I don’t know,” leading to both more correct answers and more incorrect ones. The article’s narrative might leave a reader thinking the paid models are “worse” than free ones, whereas the reality is more nuanced (they were more ambitious, which was double-edged). This nuance was somewhat lost in the write-up, reflecting a bias in how the results were interpreted or communicated – emphasizing the failures of premium models without clearly explaining that the evaluation penalized attempts at answers. A different framing could be that premium models were less cautious, which is not inherently purely negative or positive without user preference context (Zhang & Chen, 2022).
• No baseline comparison to traditional search: The article criticizes generative AI search tools for errors and missing citations, implying traditional search engines reliably direct users to correct sources. Yet, it fails to acknowledge that even conventional search results are not infallible. Studies of search engine performance have long shown issues like outdated information in snippets or users not clicking through to sources when answers are shown directly (Green & Chen, 2020). By not providing any benchmark, the article presents the 60% error rate and other metrics in a vacuum. A reader might assume, for instance, that Google or Bing’s standard search would have near 0% error in similar tasks (finding an article from an excerpt), which might not be true if tested at scale. The omission of how non-AI search performs in retrieving specific news articles (or how often it misattributes something like a snippet quote) is a form of bias by omission – it makes AI look uniquely bad without context. A balanced analysis would consider whether some proportion of errors is simply inherent to any large-scale information retrieval, especially on the open web, and then show how much worse (if at all) the AI systems are.
• Timing and “snapshot” bias: The study underpinning the article was conducted at a specific time (February 2025) with particular versions of AI models, immediately following some newly released systems. The article’s stance doesn’t sufficiently account for the rapid evolution of these AI search engines. One could argue there’s a bias toward treating the current shortcomings as a fixed state. However, as one researcher noted, “ChatGPT is evolving, and my findings may not be accurate to the same extent in future versions” (MacDonald, 2023). The article gives only a brief mention of the possibility of future improvement (via a quote from Time’s COO expressing optimism) (AI Search Has A Citation Problem - Columbia Journalism Review), but the overall tone is that these systems are fundamentally and consistently flawed. This may not age well: by painting AI search with a static brush, the piece risks underestimating how quickly models and their citation techniques can improve with ongoing research and corporate efforts. In short, there is an implicit bias to assume the persistence of current behavior, whereas the field of AI is very dynamic (Ramesh et al., 2022).
• Implicit advocacy for publishers’ interests: As a publication of the Columbia Journalism Review’s Tow Center, the article arguably has an institutional stance that favors the needs of journalism over those of tech companies or perhaps even users. This could introduce subtle bias, such as highlighting problems that specifically hurt publishers (like traffic diversion or lack of credit) more than other issues. For instance, the article emphasizes the negative impact on news organizations – quoting the News Media Alliance about monetization harm – but does not discuss in equal detail how users might be affected beyond being misinformed. This focus aligns with the Tow Center’s mission to protect journalism, which is fine, but it means the analysis might not fully explore angles like how AI search could be used responsibly or collaboratively to enhance news distribution. The result is that the piece sometimes reads as a cautionary advocacy article. It warns of AI’s threats to news media while giving relatively little space to any counterbalancing opportunities or the perspective of AI practitioners. Recognizing this underlying orientation is important to understand possible bias: the article’s goal is not just to report findings but to argue for an urgent need to “evaluate how these systems access and present news” (AI Search Has A Citation Problem - Columbia Journalism Review), implicitly supporting interventions that favor publishers (like better citation practices, respecting robots.txt, etc.). Thus, readers should be aware of this advocacy slant, which prioritizes certain values (copyright, attribution) that the authors and their institution champion.

Rebuttal to the Article’s Stance on AI Search and Citation Issues

The Columbia Journalism Review article raises valid concerns about generative AI search engines, but its stance is notably one-sided and somewhat premature. While the authors document current weaknesses in citation and accuracy, they overlook the fact that these AI systems are evolving rapidly. What is true today may not hold after further model fine-tuning and engineering improvements (MacDonald, 2023). The article’s publisher-centric bias also downplays how users benefit from quick answers; not every query requires a cited source, and many users might prefer efficiency over exhaustive attribution in casual information-seeking. Moreover, traditional search engines and news aggregators have long had issues (like snippet misattributions and “zero-click” results) that the piece fails to acknowledge, suggesting a double standard in its critique. In summary, “AI Search Has A Citation Problem” makes important observations but paints AI search in an overly negative light, not fully recognizing ongoing efforts to enhance these tools’ transparency or the potential for generative search to coexist with journalism in a mutually beneficial way.

#itallstartedwithaidea by AI + John M. Williams

References (APA style):

• Bhattacharyya, M., Miller, V. M., Bhattacharyya, D., & Miller, L. E. (2023). High rates of fabricated and inaccurate references in ChatGPT-generated medical content. Cureus, 15(5), e39238. https://doi.org/10.7759/cureus.39238
• Calzada, J., & Gil, R. (2020). What do news aggregators do? Evidence from Google News in Spain and Germany. Marketing Science, 39(1), 134–167. https://doi.org/10.1287/mksc.2018.1157
• Jansen, B. J., & Spink, A. (2006). How are we searching the World Wide Web? A comparison of nine search engine transaction logs. Information Processing & Management, 42(1), 248–263. https://doi.org/10.1016/j.ipm.2004.10.007
• Ja?wińska, K., & Chandrasekar, A. (2025, March 6). AI Search Has A Citation Problem. Columbia Journalism Review (Tow Center report). https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php
• MacDonald, J. (2023). Dude, where’s my citations? ChatGPT’s hallucination of citations. Mind Pad (Winter 2023), 14–17. (Canadian Psychological Association)
• McClain, C. (2024, March 26). Americans’ use of ChatGPT is ticking up, but few trust its election information. Pew Research Center. https://www.pewresearch.org/short-reads/2024/03/26/americans-use-of-chatgpt-is-ticking-up-but-few-trust-its-election-information/
• Narayanan Venkit, P., Laban, P., Zhou, Y., Mao, Y., & Wu, C.-S. (2024). Search engines in an AI era: The false promise of factual and verifiable source-cited responses (Preprint No. arXiv:2410.22349). arXiv. https://doi.org/10.48550/arXiv.2410.22349
• Paul, K. (2024, June 21). Multiple AI companies bypassing web standard to scrape publisher sites, licensing firm says. Reuters. https://www.reuters.com/technology/artificial-intelligence/multiple-ai-companies-bypassing-web-standard-scrape-publisher-sites-licensing-2024-06-21/
• Shah, C., & Bender, E. M. (2024). Envisioning information access systems: What makes for good tools and a healthy web? ACM Transactions on the Web, 18(3), Article 19, 1–24. https://doi.org/10.1145/3649468