21-05-2025
What AI detection tools got wrong in the case of a photo tweeted by a French politician
"I bet a 100 bucks that this is AI,' reads a comment on a May 11 tweet by the three-time French presidential candidate and far-left politician Jean-Luc Mélenchon.
The comment, left by an account named Lapin du Futur (Futuristic Rabbit), accuses the leader of the French far-left political party France Unbowed (LFI) of having used artificial intelligence to modify a photo he shared of an anti-Islamophobia protest that took place the same day in Paris.
One of the photos tweeted by Mélenchon shows him standing with several deputies from his party on the front lines of a protest. Behind them, there are thousands of protesters brandishing many different flags, including the French flag, the Palestinian flag and LFI's flag.
A number of social media users, like Lapin du Futur, claimed that the photo was generated by AI and that the French flags were added.
Lapin du Futur's post alone garnered more than two million views.
"The prompt [Editor's note: the request provided to AI software] was 'add French flags'. None of the French flags are real,' claimed another social media account in a post that garnered more than 700,000 views.
Some social media users have pointed to certain details in the photo that often indicate that an image has been AI-generated, like people's hands looking strange or deformed. In the case of this photo, one protester holding up a flag appears to have only four fingers. Another person seems to have six fingers.
Another social media user shared screenshots from an AI-detection tool, which determined that it was 'likely' that the image was generated by AI.
AI-detection tool Sightengine determined that there was a 90% probability that the image was generated by AI. Another tool, Decopy AI, said there was a 95% probability.
And yet, there were French flags
However, this image was not AI-generated and the French flags were not added to the image. No other elements were either.
A number of LFI deputies also took to social media to post photos taken during the protest on Place de la Bastille in Paris, including Thomas Portes and Aly Diouara. Other accounts with links to France Unbowed also shared images. In all of the photos, you can see several French flags held by protesters.
You can also see the French flags in videos of the protest posted by French media outlet BFM and images published by Agence France-Presse (AFP) on their website and YouTube page.
Some social media users have also raised questions about the number of protesters in the photo, claiming that the numbers have been artificially inflated. However, the photos by other sources show the same crowd sizes on Place de la Bastille. The Paris police prefecture counted 3,700 protesters, while protest organisers said that 15,000 people had attended.
As for the hands that appear to have four or six fingers, a close verification of the image reveals that each hand does actually have five fingers. If you look closely at the photo, then you can see the fifth finger of the hand holding the flagpole, partially hidden by the pole. As for the other hand, what looks like a sixth finger is actually an effect created by the angle of the hand and the shadow cast on the person's wrist (see more details below).
False positives detected from minor edits to contrast and colours
The FRANCE 24 Observers team contacted the press service of France Unbowed (LFI), who refuted all accusations that the image had been AI-generated. However, the party did clarify that 'contrast was added, the brightness was reduced and the vibrancy of the photo was accentuated' during the editing process.
It turns out that basic edits like this can actually confuse tools meant to detect AI-generated images like Sightengine, which was cited by a number of social media users who said Mélenchon's photo was false. Our team contacted Sightengine, who confirmed that their tool detected both 'images completely generated by AI but also real photos that contained elements generated or modified by AI' like some of the above edits.
"It might flag partial modifications or small edits or improvements carried out using generative AI tools,' said representatives of Sightengine.
France Unbowed sent the original image – without any changes to contrast or colours – to our team. When we ran it through Sightengine, it didn't detect any possible use of AI. This supports the theory that the tool initially concluded that the photo was likely AI-generated only because of the colour and contrast correction.
The same is true for the tool which initially concluded that there was a 99.13% probability that the photo posted by Mélenchon was AI-generated. In an email to our team sent on May 15, the company behind the tool said that its product 'may have some errors in the accuracy of multi-person photo recognition', indicating that it was primarily a tool 'aimed at the recognition of single-person photos'. The company promised that they would 'make immediate improvements' to the tool, following this error.
'There is AI integrated into almost all photo editing tools'
"Today, there are elements of generative AI integrated into nearly all of the [photo editing] tools that we use,' said Emmanuelle Saliba, who runs the investigation bureau at GetReal, a company that detects deepfakes. The company was founded by Hany Farid, an expert in the analysis of manipulated photos and videos. Saliba explained that Photoshop, for example, "uses AI in the 'enhance' function as well as in editing".
That's also the case for Lightroom, the photo editing programme that was used by France Unbowed. Lightroom explains on its site that it uses AI to improve the sharpness of images or to adjust " colours and tones'.
After a careful analysis of the image using both detection algorithms and analysis of the shadows, GetReal concluded that the photo posted by Mélenchon is 'real'.
A number of other AI-detection tools also did not conclude that the flags were AI-generated, including the algorithms developed by a European research and development project focused on disinformation analysis.
"None of these algorithms shows any sign of AI generation in these images,' said Denis Teyssou, the editorial lead at AFP's Medialab and
"We are looking for something tangible, especially signs left by image generators in the signal,' he said, unlike other AI detectors, which might say that an image is likely generated by AI because of small edits on colour or contrast.
Due to the proliferation of audio, video, and images generated by AI, numerous free AI-detection tools are now available. However, it is important to remain prudent as these tools can produce false positives, wrongly identifying content as artificially generated. This is a major challenge as most of these tools do not explain their criteria for detection, nor the technical aspects that led to images being detected as AI-generated.
