Donald Trump putting his hand on the thigh of an underage girl, Kamala Harris screaming and spitting blood: Eight weeks before the US presidential election, artificially generated images of the candidates are circulating on social media. None of them bear any kind of "AI-generated" label, which would make it easier to distinguish them from the real photos posted online.

Meta, the parent company of Facebook and Instagram, said it intended to add this type of warning. In a press release published in February, it declared that it was working on an "AI generated" signal with a view to the numerous major democratic events of 2024. But six months on, little progress has been made: For the time being, there are no warnings of this kind on Facebook, only on Instagram (the social network's internet site is still devoid of them).

These discreet labels take the form of an "AI info" note that users must click on to get an explanation, but they are still relatively rare. Le Monde found that most images on Instagram representing American political figures do not have the label, even when they harm their public image. For example, doctored images of Harris in front of a communist audience at a DNC in Chicago or on the arm of sexual predator Jeffrey Epstein have not been systematically labeled.

Meta's computers simply look for a mark (visible or invisible) that various generative services automatically embed in images to facilitate their identification. In its press release, Meta stated that it is working on labeling images with the most high-profile services, operated by Midjourney, OpenAI, Microsoft, Adobe and Google in particular. This excludes many smaller image generators.

But these loopholes in Meta's approach don't explain everything. What justification is there for the absence of an AI label on most images generated by Midjourney, with which Meta works? Perhaps the fault lies in the difficulty of correctly positioning the "cursor" that separates AI-generated content from other content. In June, Meta came up against the rage of photographers whose images had been mistakenly tagged. The American company did not respond to requests for comments on the issue.

To fill these gaps, the solution could come from an automatic detection system that spots AI-generated images, even if they are not labeled as such. This is what various independent services claim to be able to do, including Hive Moderation, which states that it is can spot the creations of a score of AI-generation tools. Meta is working on its own detection system, but the group told Le Monde that it is not yet operational.

You have 54.63% of this article left to read. The rest is for subscribers only.