AI at Frontières: 37 out of 50 articles exceed 50% according to Pangram, a result that needs to be qualified
On September 29, 2026, CheckNews reported that 37 out of 50 Frontières articles submitted to the Pangram detector received an AI usage score exceeding 50%. This result is a signal to be examined, not a direct indicator of the media outlet’s editorial practices. Our verification therefore distinguishes the reported figure from the broader conclusion that Frontières makes extensive use of AI for articles on its website. ( collector.com.tr )
The claim has been verified
The figure cited comes from CheckNews: among 50 articles published on the Frontières website and tested with Pangram, 37 reportedly exceeded the 50% threshold. The question posed in the title of the Libération article is more ambitious: do these results establish that Frontières makes extensive use of AI? The available information does not contain a statement attributed to Frontières. The article is dated September 29, 2026; the publication period of the 50 tested texts is not specified in the available summary. ( collector.com.tr )
What the facts show
According to figures published by CheckNews, 37 out of 50 represents 74% of the articles tested . This percentage describes the stated sample, and not necessarily all of Frontières’ articles. It also does not mean that « 74% of the site » was : the result refers to the number of articles that crossed a threshold set for each text. ( collector.com.tr )
The score also needs to be interpreted correctly. Pangram presents a 50% score as an estimate of the proportion of the document it classifies as or AI-assisted . This is neither proof that exactly half the words were produced by a chatbot, nor a 50% probability that the author used AI. The Pangram 4 documentation, in fact, distinguishes between segments « written by a human, » « AI-assisted, » and « . » ( pangram.com )
The verification
A detector is not without merit, but its performance must be considered in relation to what it has actually been tested on. A study published in June 2026 compared Pangram to three other tools on 160 English-language academic texts of known origin. Pangram achieved the best results of the tools studied. However, the researchers specify that this experiment, conducted on long, academic texts, does not reproduce all real-world writing conditions; their second corpus, composed of authentic theses, did not have a verified origin to control each detection. These results therefore do not automatically validate a score applied to a particular French press article. ( link.springer.com )
Another crucial point for reproducing the test is the Pangram version . The provider indicates that Pangram 4 has been available since July 29, 2026, but that version 3.3.2 remained the default version in its programming interface until September 30, 2026, unless version 4 was explicitly selected. Therefore, the results of a new run through the tool would not necessarily be comparable without knowing the version used by CheckNews. ( pangram.com )
We were unable to access the full text of Libération’s investigation or reproduce its 50 analyses. The available information does not include a list of articles, their dates, the submitted texts, individual scores, the settings used, or the sample selection method. Therefore, we cannot independently audit the count of 37 articles or determine whether these 50 texts represent the site’s output. This limitation does not prove that CheckNews’s result is erroneous; it simply prevents us from presenting it here as a finding that has been verified a second time.
Sources and data
This verification is based on the numerical claim published by CheckNews, on technical documentation and explanations from Pangram regarding its scores and versions, as well as on an independent scientific evaluation of the detector. These sources shed light on what the test might mean. They do not, on their own, provide the complete history of Frontières’ articles. ( collector.com.tr )
Verdict
Verdict: impossible to confirm. The figure of 37 out of 50 articles is a result reported by CheckNews , which we were unable to recalculate. Even if it were replicated, a detection score would still be an indicator of or AI-assisted content, and not a direct observation of Frontières’ working methods. To assert without this caveat that the media outlet « heavily » uses AI across its entire site would go beyond what can be verified here.
Key takeaways
The issue warrants investigation; a software score alone does not replace knowledge of the texts tested, the sampling method, and the writing process. The distinction is simple: 37 reported detections in a sample are not the same as 37 established writing processes .
Main source: Libération (www.liberation.fr)
Original article: See the original source
Author: Jacques Pezet
Sources of verification
