Study Reveals Over 60% of AI Chatbot Responses Are Incorrect

AI Chatbots and the Challenge of Accurate News Attribution

A recent study published by the Columbia Journalism Review’s Tow Center for Digital Journalism reveals that commonly used AI chatbots provide incorrect or misleading information over 60% of the time. The research raises alarms about these tools potentially eroding public trust in journalism and adversely affecting the revenue and traffic of news publishers.

Exposing the Flaws of AI Chatbots

Researchers evaluated eight different generative AI chatbots—including ChatGPT, Perplexity, Gemini, and Grok—to determine their ability to accurately identify sources for 200 excerpts from recent news articles. The results highlighted serious issues: more than 60% of the chatbot responses were found to be inaccurate. Common faults included:

• Invented headlines that never existed.
• Failure to attribute articles correctly.
• Citing unauthorized copies of original content.
• Linking to broken URLs or unrelated pages even when accurate sources were named.

Reassuring but Misleading Confidence

One of the striking findings was that these chatbots often provided incorrect answers with unwarranted confidence. For instance, ChatGPT showed inaccuracies in 134 out of 200 queries but only indicated doubt on 15 occasions. Interestingly, premium models, such as Perplexity Pro and Grok 3, were found to perform worse than free alternatives when it came to accuracy, providing more definitively incorrect answers.

Moreover, five of the evaluated chatbots bypassed efforts by publishers to restrict access to their information. For instance, Perplexity cited articles from National Geographic while still being hindered by their own crawlers, and ChatGPT often relied on unauthorized reposts to access paywalled content.

The Syndication Dilemma

Many AI chatbots redirected users to syndicated articles hosted on platforms like AOL or Yahoo instead of the original news sources—even when content licensing agreements were in place. This raises concerns about revenue loss for publishers. For instance, Perplexity Pro cited entries from the Texas Tribune without providing proper attribution, despite a partnership between the two. Additionally, Grok 3 and Gemini frequently generated incorrect URLs, leading users to error pages.

While companies like OpenAI and Perplexity have entered into agreements with publishers like The Guardian and Time, the accuracy of these citations remains questionable. One study showed that ChatGPT only correctly identified one out of ten excerpts from partnered outlets.

The Impact on News Publishers

The implications for news organizations are significant. AI tools are increasingly replacing traditional search engines, with a notable percentage of the population in the U.S. now relying on them for information. Unlike Google, which typically drives traffic back to news sites, these chatbots summarize content without offering links back, thereby restricting publishers from earning ad revenue.

Misidentifications can further tarnish the reputations of these publishers. Credible brands, when misrepresented in chatbot responses, can inadvertently lose the trust of the public, endangering the overall perception of media integrity.

Future Considerations

When approached about the study, representatives from OpenAI and Microsoft defended their practices but did not specifically address the findings. OpenAI emphasized its commitment to respecting publisher preferences, while Microsoft stated that it adheres to ‘robots.txt’ guidelines. Researchers argue for systemic overhaul, calling for AI companies to enhance transparency, accuracy, and respect for publishers’ rights.

Despite existing challenges, there are indications of hope. Experts suggest that improvements in AI technologies are inevitable. However, they caution users to be wary of relying too heavily on free products that may not always be trustworthy.

As the situation evolves, it is clear that the AI landscape is changing how audiences access information and poses a growing challenge for news publishers to protect both their content and integrity.

For further insights, the detailed research can be explored here.