A few days ago, a new report once again shook the already shaky foundations of confidence in generative artificial intelligence. A international study conducted in the public media in 22 countries revealed a disturbing fact: 45% of the responses provided by leading AI assistants (ChatGPT, Copilot, Gemini and Perplexity) by leading AI assistants (ChatGPT, Copilot, Gemini and Perplexity) contained significant errors. contained significant errors.

The problem is particularly relevant in a scenario where these tools coexist with (and in some cases replace) traditional search engines. According to data reported in October 2025ChatGPT reaches 3.2 billion monthly users.

Its exponential growth (2 billion in March) is not slowing down other platforms either. Gemini could reach 450 million users, but its growth could be even greater, as Google is integrating Gemini into all its apps and operating systems.

Beyond the figures, the truth is that many people today take for granted the results of their consultations with these assistants. And they may be making serious mistakes. The study reports that the 81% of the responses analyzed presented some type of problem.

Around 33% of the citations were absent, misattributed or altered. Thirteen percent of the citations are invented or modified, and one of the now classic drawbacks of the network persists. Many of the citations are out of date and much of the data contains obsolete information.

Added to this are other challenges for which AI does not yet seem to have an answer. The confusion between opinion and facts or the loss of context. The omission of essential nuances in legal, health or political issues can generate half-baked answers that generate confusion or elude relevant points.

These are some of the conclusions reached by a group of professional journalists who analyzed 3,000 responses, assessing factual accuracy; the reliability of sources and quotes; and the ability to distinguish fact from opinion. The research, led by the BBC and involving public media such as RTVE, covered 14 languages and 22 countries, which made it possible to observe transversal patterns of error.

The analysis shows that these failures are not isolated incidents, but cross-border, multilingual and systemic. In other words, they are common to the millions of users who now also use artificial intelligence to get news.

As Jean Philip De Tender points outThe problem with these distortions is that they can jeopardize public trust, said the director of media at the European Broadcasting Union (EBU). “When people don’t know what to trust, they end up not trusting anything, and that can deter democratic participation,” he said at the presentation of the study.

For his part, David Corral, head of International Relations and Cooperation at Radiotelevisión Española, insists on the seriousness of the distortion generated since, in many cases, the results of the consultations do not include references to the original source. “If we take for granted what the attendees tell us, the media and the institutions may cease to have value for society”, he warns, in exclusive statements to DigitalES.

Corral, who values the work developed by his colleagues at TVE in this study, recalls the need to guarantee a public service such as access to accurate information. “If the information is diluted or not entirely correct, our work as media can be misinterpreted,” he denounces.

In this line, he points out that AI assistants have become new intermediaries between the media and citizens. For this reason, he advocates a better relationship between the big tech companies and the media that will help everyone to better understand the news content.

Where to intervene and what measures to take

The solution is far from demonizing generative artificial intelligence. It is more than proven (hence its rapid adoption), its ability to create original content – such as text, images, music or code – quickly and efficiently, saving time and resources.

IAG facilitates the automation of creative and repetitive tasks, drives innovation in sectors such as education, design, medicine or engineering, and improves the personalization of products and services. It also helps in decision making by analyzing large volumes of data and generating ideas or solutions that extend human capabilities.

We cannot give up on the range of possibilities that the use of artificial intelligence opens up, and that is why the authors of the study propose a series of indications for solving the problems that have been identified in the responses generated by AI assistants.

If there is one thing that this technology has demonstrated in recent times, it is its ability to learn quickly, so it is to be expected that, by applying the right formula, the results may be tighter in the near future.

In the journalistic productionThe experts maintain that it is advisable to establish mandatory human review when AI is used in search or writing tasks. It is also advisable to implement editorial verification checklists (dates, context, quotes) and to maintain clear identification of the news genre (news, analysis, opinion).

In the technical publicationsIn technical publications, they suggest incorporating AI-readable metadata (author, date of update, type of content); applying traceability systems (C2PA, watermarks) to authenticate the source; controlling robots and tracking permissions, avoiding uses that alter citations or context.

In the distribution and relations with platformsIn the distribution of content, they propose to require that attendees cite source, link and timestamp correctly; to request freshness check mechanisms; and to encourage access to content through certified editorial APIs, not through scraping.

Finally, in the user interface and experienceFinally, in the interface and user experience, they advocate requesting that platforms implement visible disclaimers when information is unverified or outdated.

What is certain is that today the various public broadcasters involved in the report are pressing national and EU regulators to enforce existing laws on information integrity, digital services and media pluralism. They also insist that continued independent monitoring of AI assistants is essential, given the rapid pace of development of this technology.

Will it be enough to achieve a higher quality of responses from attendees? Only time will tell, but any effort in this direction seems wise so as not to undermine confidence in a technology that can bring many positive things to society as a whole.