A recent study by the Columbia Journalism Review’s Tow Center for Digital Journalism has revealed alarming accuracy issues in AI-powered search engines. The research, which tested eight generative AI search tools with live search capabilities, found that these models provided incorrect information in over 60% of queries related to news sources.
This raises significant concerns about the reliability of AI-driven search tools, especially as they become increasingly popular alternatives to traditional search engines.
The study, conducted by researchers Klaudia Jaźwińska and Aisvarya Chandrasekar, involved 1,600 queries across eight AI models. The researchers tested the tools by providing excerpts from real news articles and asking the AI to identify the article’s headline, publisher, publication date, and URL.
The results were troubling: some models, such as Grok 3, had error rates as high as 94%, while others, like Perplexity, performed slightly better with a 37% error rate. ChatGPT Search, a widely used tool, incorrectly identified 67% of the articles tested.
One of the most concerning findings was the tendency of these AI models to provide “confabulations”—plausible but incorrect answers—rather than declining to respond when they lacked reliable information. This behavior was consistent across all tested models, including premium versions like Perplexity Pro and Grok 3’s paid service. Surprisingly, the premium versions often performed worse than their free counterparts, as they were more likely to confidently deliver incorrect responses.
The study also highlighted issues with how AI search tools handle publisher content. Many models ignored Robot Exclusion Protocol settings, which are used by publishers to block unauthorized access. For instance, Perplexity’s free version successfully identified excerpts from paywalled National Geographic content, despite explicit restrictions.
Additionally, these tools frequently directed users to syndicated versions of articles on platforms like Yahoo News instead of the original publisher’s site. Fabricated URLs were another major problem, with more than half of the citations from tools like Google’s Gemini and Grok 3 leading to broken links or error pages.
AI and information quality
These issues create significant challenges for publishers, who must decide whether to block AI crawlers entirely or allow them to reuse content without driving traffic back to their websites.
Mark Howard, COO of Time magazine, expressed concerns about transparency and control over how AI tools use their content. While he acknowledged the potential for improvement in future iterations of these tools, he also criticized users for expecting free AI services to be fully accurate, stating, “If anybody as a consumer is right now believing that any of these free products are going to be 100 percent accurate, then shame on them.”
Despite the findings, companies like OpenAI and Microsoft have yet to address the specific issues raised in the study. OpenAI has promised to support publishers by providing clear attribution and links, while Microsoft claims to adhere to publisher directives and exclusion protocols. However, the study underscores the urgent need for improvements in AI search tools to ensure accuracy, transparency, and respect for intellectual property. As these tools continue to evolve, their current shortcomings highlight the importance of skepticism and critical thinking when using AI for information retrieval.
