On AI-generated images, labels to recognize them are still rare

Monday 09th September 2024 02:49 PM

Donald Trump putting his hand on a minor’s thigh, Kamala Harris screaming and spitting blood or represented as a puppet : at eight weeks of the American presidential election, images of the candidates generated by artificial intelligence (AI) are circulating on social networks. Not all of them bear an “AI-generated” label, which would make it easier to distinguish them from the real photos that are circulating.

This doctored photo of Kamala Harris was posted on August 20 on the official Facebook page of Donald Trump’s campaign, without any specific mention. The warning message here, as on the other images contained in this article, was added by “Le Monde”. SCREENSHOT “THE WORLD”

Meta, the parent company of Facebook and Instagram, had nevertheless stated its intention to add this type of warning. In a press release published in February, it declared that it was working on it in view of the major democratic events of 2024, which will be particularly numerous. But, six months later, the project has made little progress: for the time being, there is no warning of this type on Facebook, but only on the Instagram application (the social network’s website is still private).

Instagram’s very discreet “AI Info” warning appears just above the image. When clicked, the following text appears: “AI Notice. Generative AI may have been used to create or modify content in this post.” SCREENSHOT “THE WORLD”

Discreet, these labels take the form of a mention « AI info » on which the user must click to obtain explanations. They are still relatively rare: according to the findings of the Mondemost images on Instagram depicting American political figures do not wear them, even when they harm their public image. For example, doctored images of Kamala Harris in front of a communist assembly or on the arm of sexual predator Jeffrey Epstein are far from being systematically stamped.

Because the system Meta uses to identify these images is not perfect. For now, its computers are content to look for a marking (visible or invisible) that various generation services automatically embed in images in order to facilitate their identification. In its press release, Meta states that it is working on this marking with the most media-friendly services, operated by Midjourney, OpenAI, Microsoft, Adobe and Google in particular. Which excludes a number of small image generators.

But these holes in Meta’s racket don’t explain everything. How else can we explain the absence of labels on most of the images generated by Midjourney, with which Meta works? The fault may lie, perhaps, in the difficulty of correctly positioning the “slider” used to distinguish between AI-generated content and others. In June, Meta faced the anger of photographers whose images had been labeled by mistake. When asked about this, the American group declined to respond.

You have 61.06% of this article left to read. The rest is reserved for subscribers.