Imagine that you ask your AI assistant about recent information that you saw in a means of your trust, such as ELECONOMIST. It responds safely but citing another medium, another holder and a non -existent URL. It is not an exception, but a usual error.
According to a study of Tow Center for Digital Journalismgenerative search tools such as Chatgpt, Perplexity, Grok either Geminicandidates for unseeing to the classical search in Google and other alternatives for their greater comfort, are not only wrong to quote news, but do it with a conviction that can induce the error even to the most attentive user.
The rise of these IAS as substitutes for traditional search engines has been meteoric. But its growth raises an awkward question: What happens when these tools not only fail to quote correctly, but also ignore the rules of the game imposed by the means themselves? Are we facing a new form of digital extractivism that empties traffic to the original sources?
6 out of 10 answers with AI are wrong
The study, published on March 6, 2025, analyzes the performance of eight generative search engines in its ability to properly cite journalistic content. The investigation focused on assessing whether these models were able to correctly identify the holder, the medium, the publication date and the original URL of informative items, from fragments extracted directly from the pieces.
The findings show a worrying panorama: more of the 60 % of the answers were incorrectand in many cases, generative tools responded with unjustified confidence. In addition, several models ignored the guidelines for the exclusion of trackers (robots.txt), accessing restricted content or even after payment wall. Even among those with license agreements with media, such as Perplexity and OpenAIno significant improvement was observed in the accuracy of appointments.
The study also shows the tendency to MANUFACTE ERRONE URL, attribute content to erroneous media and Replace original sources with syndicated versionsall without recognition of the error or explicit decline of response.
The ranking of generative search engines according to their reliability
The study compared as we say eight platforms. Each one was performed 200 consultationsmeasuring the ability to correctly cite the holder, the medium and the URL of the origin article.
This is the ranking and its global performance in terms of answers completely incorrect:
- Perplexity (free) – 37 % of incorrect answers
- Chatgpt – 67 % of incorrect answers
- COPILOT – Approximately 70 % of incorrect answers (High non -answers rate)
- Perplexity Pro – Approximately 72 % of incorrect answers
- Gemini – 77 % of incorrect answers
- Deepseek – 80 % of incorrect answers
- Grok 2 – 90 % of incorrect answers
- Grok 3 – 94 % of incorrect answers
Although Perplexity is free It was the one that was least wrong, it also incurred bad practices such as accessing restricted content and mentioning syndicated versions instead of the originals. For its part, Grok 3 demonstrated an almost total inability to offer precise answers and also led in the generation of False or broken urls.
The illusion of certainty: erroneous but confident responses
One of the most disturbing findings in the study was the attitude with which the chatbots were wrong. Instead of recognizing their limits, they gave categorical responses, without leaving space for doubt. The use of expressions such as “could be” or “I don’t have enough information” was rare.
Chatgptfor example, he only expressed insecurity in 15 of the 200 incorrect responses. In no case refused to answer. This pattern was repeated in almost all tools except COPILOTwhich was the only one who declined more answers than those that respondedprobably due to its integration with Bingbot, which allows you to more easily respect the media content.
This unfounded trust is dangerous. As the study authors point out, the authoritarian tone of the answers can make users valid completely false or poorly attributed information. And it is not just a technical error: it is a matter of Information Ethics.
The attribution problem: when the media disappear
Another common mistake was the Incorrect attribution of contents. Many models, by misunderstanding the environment or not providing the original link, ended up redirecting to syndicated versions (for example, in Yahoo News or AOL), which deprives the means of traffic and visibility.
This phenomenon is not accidental. Even when there were trade agreements between means and the Texas Tribune With Perplexity, the tools chose to cite unofficial copies. Thus, the basic principle of the web is broken: Who publishes, receives traffic.
Even worse: even means that actively blocked the access of certain crawlers or trackers (such as National Geographic either The New York Times) They saw their content appear in the results, citing unauthorized sources. This indicates that some IAS, as perplexitythey could be avoiding or ignoring the robot exclusion protocol, which puts the editorial control model in check.
Although OpenAi and Perplexity lead in the signing of agreements with media (including names such as Time, The Guardian either Schibsted), the study reveals that these alliances do not translate into greater reliability. In many cases, the contents of the partners were summoned with the same erroneous frequency as those of media without a contractual relationship.
For example, Timeboth Openai and Perplexity partner, was correctly cited more than others, but never consistently. Instead, the San Francisco Chroniclepart of the Hearst group and with access to the OpenAi Crawler, it was correctly cited only Once ten possible. Not even in that single case a valid link was included.
These failures feed a legitimate suspicion: Are these agreements for something more than to improve the public image of the platforms?
#trust #artificial #intelligence #paying #study #reveals #appointments #wrong
