As it becomes more and more difficult to distinguish an image generated by artificial intelligence with the naked eye, many online detection tools claim to be able to detect them. But a test carried out by the France 24 Observers editorial team shows why their results must be interpreted with caution.
Published on: 04/16/2026 - 09:37
How to spot an AI-generated image? Faced with the constant improvement of image generation software, the detection reflexes that we could still mobilize at the start of the development of these technologies - spotting poorly made hands or incoherent faces - already seem outdated.
Many Internet users are now turning to detection tools. Free and easily accessible online, many of them promise “reliable” and “precise” results. But these sometimes turn out to be incorrect, or even misleading, when tested on AI-generated images, but also on real images.
Also read: Images and artificial intelligence: how to spot misinformation
“False negatives”
The editorial staff of France 24 Observers tested the first ten detectors found online by typing "AI image detectors" on Google, on either real or false images. Each time, several detectors got their results wrong.
This is the case, for example, of the two AI-generated images used for this test, one supposed to show Nicolas Maduro and the other Jeffrey Epstein. These two visuals, which have gone viral in recent months and identified as AI-generated by many verification media (here and here), were identified as AI-generated by less than half of the detectors tested.
The first depicted Venezuelan President Nicolas Maduro, captured on January 3 by the United States, exiting a plane surrounded by two American soldiers. Our test shows that six out of ten tools, or more than half, incorrectly declared this image to be real. Among these bad results, some declare that the visual is “100% human” (Isgen.ai), or that it is a “99.99%” “human photograph” (Mydetector.AI).
Same thing for an image supposed to show the child molester Jeffrey Epstein still alive in Tel Aviv in 2026 (he died in custody in August 2019): this time, seven detectors did not establish that it had been generated by AI.
“Developing AI detection software is complex”
These cases where the tools fail to identify an AI-generated image - they are called "false negatives" - can be explained by the very way in which these tools work and are "trained": "the majority of online AI detectors work like classifiers: they determine whether an image is natural or synthetic", explains Tina Nikoukhah, vice-president of Research at GetReal Security, a cybersecurity company specializing in verifying the authenticity of people and digital content.
Drawing on machine learning algorithms, these tools are trained on massive databases composed of both real images and AI-generated images. But from then on, the quality of these tools depends a lot on the training databases: “the development of AI detection software is complex and sometimes, the AI detectors have not been trained on the type of images that we want to analyze,” explains Tina Nikoukhah.
The specialist highlights, for example, the diversity of AI-generated image generation tools - from Google's Nano Banana tool to ChatGPT's Dall-E, which each produce specific types of images. “Depending on their training, tools may, for example, be effective in detecting images created by a particular AI generator, but not by another.”
The life cycle of online images also plays a determining role. Often resized, modified, compressed, these images gradually lose certain signals used by the detectors. Several companies behind these tools, such as ZeroGPT or Rephrasy, have confirmed this to our editorial staff. The same false image of Jeffrey Epstein living in 2026 in a wider shot is, for example, better detected by certain detectors.
Finally, it is sometimes difficult to determine what these tools which classify between True and False have really "learned" through their training: "In reality, we sometimes do not know why a tool detected an image as true or false", continues Tina Nikoukhah. For example, if I train my tool on a lot of AI-generated photos of flags, maybe the tool will be biased and more easily classify photos of flags as fake."
Real photos wrongly detected as AI
For Tina Nikoukhah, these tools must in reality be part of a broader detection system, based on several algorithms capable of adapting to different situations. “It’s like in medicine: to diagnose an illness, we don’t rely on a single exam, but on several techniques, from MRI to CT.”
Several of the tools surveyed acknowledge that they cannot detect all AI-generated images. “Our model can sometimes be likely to classify an image generated by AI as coming from a human”, indicates for example the company Isgen.IA, contacted by our editorial staff, whose tool did not detect either of the two images above. “But in return, we obtain significantly greater precision on human images. This allows us to avoid misclassifying human works, which could have even more serious consequences.”
Because several of the tools tested can also produce what are called "false positives", that is to say results which falsely assert that a real image is generated by AI. Our editorial staff did the test with several photos published online by political figures. A photo of Emmanuel Macron published during a meeting of heads of state in December 2025 was, for example, detected by three tools as being generated by AI.
Four tools also judged a photo published in October 2025 by the now mayor of Paris, Emmanuel Grégoire, to be false. For the former candidate for the March 2026 municipal elections in Paris Rachida Dati, there are even six tools which wrongly claimed that a photo published on her X account around the same period was generated by AI.
None of these false positives had any impact on the public debate. But they can pose a problem: because of two bad results from detection tools which had wrongly detected AI in a photo published by Jean-Luc Mélenchon, the head of La France insoumise (LFI) was accused of having added French flags to a photo showing a demonstration against Islamophobia in May 2025. The photo, which was very real, had only been retouched on contrasts and brightness by the Adobe retouching tool, which is increasingly using AI for this type of modification.
The company Sightengine, at the origin of one of the false positives, had at the time indicated that its tool could detect both "images entirely generated by AI, but also real photos which include elements generated or modified by AI" - without however specifying this nuance in the results presented to a user who would do the test.
Also read: Why Jean-Luc Mélenchon is wrongly accused of having published an AI-generated image
Combine these results with other methods
Beyond these sometimes false results, these detectors are also criticized for their lack of transparency on the results given by these detectors and the reasons behind their decisions. One of the tools questioned, Rephrasy, tells us that our test will push the company to better contextualize the results, in particular to specify “when a result risks not being reliable”.
Like ZeroGPT, several of these tools also believe that it is "preferable to combine detector results with source verification, provenance verification where possible, reverse image search and human editorial review."
Author of a guide on AI-generated image detection methods in September 2025 and the originator of the ImageWhisperer tool, which offers both AI detection and contextualization of the image and its origin online, Henk Van Ess also believes with the editorial staff of the Observers that "a detector will not solve this type of problem by itself, just as a journalist will have difficulty detecting an image without technical knowledge. The objective is to ensure that these two approaches are combined.
Keywords associated with the article
The content you are trying to access does not exist or is no longer available.




