AI Search Engines Are Failing: 60% Error Rate Exposed | Image Source: arstechnica.com
NEW YORK, March 13, 2025 - A recent study by the Columbia Journalism’s Tow Center for Digital Journalism revealed a major flaw in AI-driven search engines: creeping inaccuracy. According to the research, generic IA models designed for news searches provided incorrect responses in more than 60% of consultations, raising serious concerns about their reliability.
The study, conducted by researchers Klaudia Jáхwińska and Aisvarya Chandrasekar, tested eight AI-led research tools with real research capabilities. Among the most disturbing results, some models, including ChatGPT Search and Grok 3 of xAI, provided inaccurate responses at alarming rates. With AI-driven tools that increasingly replace traditional search engines for many users, this study highlights the risks associated with misinformation generated by these models.
What is the accuracy of AI power search engines?
To assess the accuracy of AI-based research models, researchers performed 1,600 tests in eight different tools. They provided direct excerpts of press articles and asked AI models to retrieve essential details such as the article holder, the original publisher, the publication date and the URL.
The results were:
-
- Grok 3 had the highest error rate, incorrectly answering 94% of queries.
- ChatGPT Search answered 67% of queries incorrectly.
- Perplexity was the most accurate but still provided wrong answers 37% of the time.
These results highlight the fundamental problem of AI search engines: they often generate “wrong” responses rather than admit uncertainty. This overinsurance can deceive users who trust the information generated by AI without referring to it.
Why do AI models fight news searches?
The study found that AI models tend to generate responses even when they lack reliable information. Instead of refusing to respond, they often produce plausible but incorrect details. This phenomenon, commonly known as hallucination, is a known problem in generic IA, but is particularly worrying in the context of information searches.
In addition, researchers found that premium versions of IA research tools, such as Perplexity Pro ($20/month) and Grok $3 ($40/month), were actually worse in some cases. Although these paid models provided more correct responses in general, they also showed a greater tendency to generate incorrect responses rather than admit when they lacked information.
Do AI’s search engines violate publishers’ rights?
Another disturbing finding in the study was that some AI search engines seemed to ignore the robot exclusion protocol, which allows publishers to prevent AI trackers from accessing its content. The free version of Perplexity, for example, correctly identified the 10 excerpts from the content of Paywalked National Geographic, although the publisher explicitly blocked its web trackers.
According to the study:
“Sometimes, AI chatbots have recovered information from sources that explicitly prohibited the disposal of the CEW. This suggests that some AI providers may not comply with copyright restrictions.”
This raises ethical and legal questions about how AI companies collect and use data. If AI models overcome content restrictions, they may violate intellectual property rights and compromise the ability of publishers to control access to their work.
Can IA search engines provide reliable certificates?
Beyond factual inaccuracies, the study also found that many AI search engines did not adequately cite sources. Instead of providing exact URLs or editorial details, they often invented quotes or directed users to broken links.
The researchers said:
“More than half of Gemini and Grok 3’s responses include built-up or non-functional links
When AI-generated research tools cannot provide a reliable date, they erode confidence in their ability to serve as research tools. Users who rely on IA search engines for fact control or academic work may inadvertently disseminate incorrect information due to incorrect or lost powers.
What experts say about AI search engines
Industry experts have long warned against the limitations of AI-generated content. Entrepreneur Mark Cuban recently addressed this issue at the SXSW conference, saying:
“AI is never the answer. AI is the tool. Whatever your skills, you can use AI to amplify them.”
Cuba stressed that AI should not be seen as a substitute for critical thinking or independent research. Rather, it should serve as a complementary tool to improve human experience.
Similarly, researchers Chirag Shah and Emily M. Bender expressed concern about AI’s search engines, arguing that:
- Reduce transparency by obscuring how information is sourced.
- Amplify bias present in training data.
- Encourage over-reliance on AI-generated answers.
These concerns underscore the need for users to approach the research results generated by AI carefully and verify information from traditional sources.
What this means for the future of AI research
The increasing adoption of AI-based research tools, particularly among young users, raises serious questions about the future of access to information. If the search engines generated by AI continue to produce inaccurate or misleading results, users can develop a distorted understanding of current events and historical facts.
The experts recommend a number of measures to mitigate these risks:
- AI developers should prioritize accuracy over engagement. Instead of optimizing for confident-sounding responses, models should be trained to acknowledge uncertainty when necessary.
- Regulators may need to enforce stricter transparency standards. AI companies should be required to disclose how their models retrieve and verify information.
- Users must develop critical thinking skills. Relying solely on AI for news and research can lead to misinformation. Cross-referencing sources remains essential.
As artificial intelligence technology continues to evolve, it will be essential to address these issues to ensure that search engines remain reliable sources of information.
The results of this study are a clear reminder: AI research tools are far from perfect. While offering comfort, they do not replace a complete and independent search. As the bad information generated by AI becomes more frequent, users must exercise caution and verify the facts before accepting them as truth.
