Search-Based AI is not ready for Prime Time : Hallucinations and Citation issues | Openai | Best course on large language models online | A compact guide to large language models pdf 2021 | Turtles AI

Search-Based AI is not ready for Prime Time : Hallucinations and Citation issues
Systematic errors in citation and access to content: the challenge of AI for the future of information
Editorial Team15 March 2025

 

The Tow Center report highlights that AI-based search tools, while increasingly popular, are failing to cite sources and retrieve accurate content, compromising the reliability of information and traffic to publishers.

Key Points:

  • High rate of errors and inaccurate citations. 
  • Marked differences between premium and free versions. 
  • Violation of publisher preferences (robots.txt). 
  • Risk to publisher traffic and visibility.


In the current context, the growing adoption of AI search engines has raised serious concerns about the correct retrieval and citation of journalistic content, as highlighted by the recent Tow Center for Digital Journalism report, which highlighted that of over 1,600 queries tested, more than 60% produced incorrect answers characterized by incorrect attributions, invented URLs, and a strong propensity to provide definitive answers without admitting uncertainty, thus generating a false impression of reliability; paradoxically, the premium versions, despite having higher costs and supposedly higher capabilities, have proven to be more prone to errors with deceptive security compared to their free counterparts, exposing a critical issue in managing access to content, as exclusion mechanisms based on robots.txt, although theoretically supposed to protect publishers, are frequently ignored or bypassed, with damaging consequences for organic traffic and the recognition of original content, a phenomenon confirmed by further analyses and reports by entities such as Nieman Lab, HD Blog and Multiplayer.it, which highlight how such gaps can compromise not only the transparency of citations but also the economic sustainability of the journalistic sector; corollary: only through a collaborative and transparent approach between AI developers and publishers will it be possible to mitigate these risks and enhance the potential of search tools.

In particular, OpenAI’s ChatGPT Search showed significant shortcomings. Out of 200 citations analyzed, 153 were partially or completely incorrect, with the chatbot attributing citations to non-existent or incorrect sources. For example, it attributed a quote from an Orlando Sentinel letter to the editor to a Time article.

Another example involved a quote from a New York Times article about endangered whales, for which ChatGPT provided a link to a site that plagiarized the original article. The chatbot also rarely admits uncertainty in its responses; in 200 tests, it only used phrases like “may” or “might” in seven cases or said it couldn’t find the exact article.

These findings raise questions about the ability of AI search tools to provide accurate and verifiable information, highlighting the need for significant improvements to ensure the trustworthiness of the information provided to users.