Fake photo of Trump kissing his assistant: how AI was used to cover its tracks

On the night of September 14 to 15, an alleged photographer of Donald Trump spread like wildfire on social networks. Viewed more than 10 million times on a single post alone, the image purports to show the American president on a golf course kissing his assistant, Nathalie Harp. Presented as a “special assistant” within the White House organization chart, Nathalie Harp was mocked by a democratic senator for his dedication and omnipresence in Donald Trump’s entourage. Recently, the American president granted a bonus of 45,000 dollars (38,664 euros) presented as an “end of year gift” to his colleague.

“Thirty years ago, a photo like this could have destroyed a politician’s career overnight. Today? Hardly enough to interrupt the news cycle,” laments a user who shared the alleged photo.

A second alleged photo sows trouble

Although it went viral, the supposed photo quickly raised doubts as to its authenticity. In the image, Nathalie’s right hand has six fingers, a characteristic defect of image generations using artificial intelligence.

Even more, if we pass the alleged photo through SynthID, an artificial image detection tool used by Open AI, the company that owns ChatGPT, the result reveals that the fake photo contains a watermark which indicates that it was produced by ChatGPT.

However, these elements did not make it possible to definitively conclude, on the date of their publications on September 15, on the inauthenticity of the kissing scene.

Indeed, another alleged photo of Donald Trump kissing Nathalie Harp had also appeared on social networks. It was a less clear version of the Trump photograph, in which the defect in the six fingers of Nathalie Harp’s right hand is not visible. Submitted to the SynthID artificial image detector, the blurred image does not return any results to establish that it includes a watermark attesting that it was generated by AI.

A new theory was then put forward by certain Internet users: the supposed blurry photo would in fact be authentic and the real original image. If the so-called clear image has indeed been processed by AI, these Internet users concede, it is to improve the quality of the original photo. However, this assumption is also false.

A first image shared by a pro-democrat influencer

To determine the exact origin of these images, the editorial staff of Les Observateurs carried out a reverse image search (see here how to do it) using the Google Lens tool.

To this end, our editorial team developed a computer program in Python to process approximately 900 search results from Google Lens. The goal of this program was to allow search results to be extracted from Google Lens and then sorted.

This research, which is based on the analysis of the anteriority of publications according to their date of publication, made it possible to trace the path of diffusion of the images and to confirm their artificial nature.

As the BBC journalist was also able to determine Shayan Sardarizadeh One of the first instances of Donald Trump’s photograph was posted on September 14. This is a miniature of a real (a vertical video in Facebook) which we find posted on this date on Facebook, at around 7:20 p.m., by Keith Edwards, an influencer who displays pro-democratic positions.

This September 14 version is different from the September 15 clean shot. If the image displays comparable sharpness, it has a darker exposure or appearance.

And here, the SynthID tool confirms that this sharp, darker image was indeed produced by the Gemini image-generating AI.

It should be noted that Keith Edwards also broadcast on September 13 another false AI-generated image of Donald Trump accompanied by Nathalie Harp. In this false image, the American president appears to be holding the hand of his assistant.

Our research then makes it possible to establish that in its blurred version, the first occurrence of the image of Donald Trump kissing Nathalie Harp was posted on the network ThreadsSeptember 14 at 9:08 p.m. This result was also found by Emmanuelle Saiba, researcher at Get Real, an IT security company.

And unlike the blurry image visible in the posts published on September 15, this blurry photo from September 14 is indeed detected by SynthID as an AI-generated image.

Finally, one of the first clear images, this time with clear colorimetry similar to that broadcast during the day of September 15, was broadcast on Threads on the morning of September 15, at 6:19 a.m. Again, the image is detected as AI-generated by the SynthID tool.

Image degradations and enhancements to hide anomalies

What can explain this succession of AI-generated image productions and the fact that some are detected by SynthID but not others? According to the professor Hany Faridfounder of the company Get Real, the most likely scenario would be that the authors of these fake photos wanted to degrade the original images generated by AI to hide their anomalies and the watermarks that could prove their artificial nature.

So in summary, the first clear image of Trump kissing Harp, published on September 14 at 7:20 p.m., was generated by AI. It was then probably degraded, again with the help of an AI, to produce the blurred image, published the same day at 9:08 p.m. This blurry image was then itself improved, again by artificial intelligence, to produce the clear image of September 15 published at 6:19 a.m.

If certain versions, in particular the blurred one, are not detected by SynthID, it is because the watermarks identifying artificial images, which this tool must identify, may have been lost through cropping and resolution degradations carried out by Internet users during their publications on social networks.