Eight search engines powered by generative AI models have shown significant problems with news sources, according to a study conducted by experts at the Center for Digital Journalism at the Columbia Journalism Review.
Image source: Christopher Burns / Unsplash.com
AI search engines get news source queries wrong more than 60% of the time. About one in four Americans now uses AI as an alternative to traditional search engines, the study’s authors say, and the significant error rate raises serious concerns about AI’s reliability. Error rates varied across platforms: Perplexity got the information wrong 37% of the time, while ChatGPT Search incorrectly identified 67% of sources. The chatbot Grok 3 had the highest error rate, at 94%.
In their testing, the researchers fed the AI services snippets of real news articles and asked each model to identify the article’s headline, source, original publisher, publication date, and URL. All the models had one thing in common: Without reliable information, they didn’t refuse to answer, but instead gave plausible-sounding but false answers. Notably, the paid models fared even worse in some ways: Perplexity Pro ($20 per month) and the premium Grok 3 ($40 per month) were more likely to get answers wrong than their free counterparts. They got more queries right, but they didn’t refuse to answer in the absence of reliable information, so their overall error rates were higher.
Image source: cjr.org
The researchers also confirmed that some AI platforms ignore directives that prohibit web crawlers from accessing resources. For example, the free Perplexity correctly identified all 10 pieces of paid content from National Geographic, even though the publication had explicitly banned the service from visiting its site. When citing sources, AI search engines often went to aggregators like Yahoo News rather than the original news sites. This happened even when the publishers and developers of these AI systems had official licensing agreements. More than half of the links in chats with Google Gemini and Grok 3 led to fictitious or broken URLs – in the case of Grok, this was 154 out of 200.
These issues present publishers with a difficult choice. Blocking web crawlers will only make AI search engines’ problems worse; opening up their sites entirely to them will cause users to stop visiting them, content with chatbot responses. Time magazine’s chief operating officer Mark Howard expressed hope that AI developers will improve the quality of their services in the future with significant investment. He urged chatbot users not to place too much trust in AI responses.
OpenAI and Microsoft have confirmed that their systems can produce such results. OpenAI has promised to support publishers by directing visitors to their sites with summaries, citations, and proper links; Microsoft has assured that its web crawlers follow publisher guidelines.