how social networks will frame AI-generated images

Meta will add an “imagined with AI” label to photorealistic images generated using Meta AI or any other tool. An indication that will arrive on all its social networks: Facebook, Instagram and Threads.

This is one of the great challenges in terms of artificial intelligence in 2024: the detection of content generated by AI. This is especially true for artificial photorealistic images. Meta must participate in meeting this challenge, especially since the company offers Imagine With Meta AI, a free image generation tool.

Of the ” signals » to detect AI-generated images posted to Facebook, Instagram or Threads

In a statement, Meta said it worked with some industry partners to create “ common technical standards that indicate that content has been created using AI. » This around the Partnership on AI (PAI), which brings together important technological players, such as Adobe, Amazon, Apple, Google, Intel, Microsoft, OpenAI and Samsung. It is these standards that will allow Meta to label the images concerned on its three social networks, Facebook, Instagram and Threads.

This addition will arrive in the coming months and gradually in all languages ​​of the three applications. The goal is to get there in the coming months as well as by the end of the year, since important elections will take place in 2025. This only works for the moment on images, but not on videos or audios. Meta says that companies that offer generators have not begun to include watermarking systems on these types of files.

On the user side, Meta wants to force them to label AI-generated content as such. In which case they will be exposed to sanctions from Meta. We do not know the nature of the planned sanctions. Furthermore, this February 6, OpenAI formalized its watermark system for images generated via its Dall-E 3 tool, C2PA. Beyond Meta’s desire to be honest with users of its social networks, the company may in the future have no choice. Legislation around the world is looking at generative artificial intelligence and some would like the content it generates to be indicated as such. This is the case for the United States, which could adopt its own technology. In the European Union, it is the AI ​​Act which could legislate on it: the law is currently being drafted and negotiated.

Meta’s image generation tool also begins to label content

Meta wants to show off, including with its own tool, currently in beta (and unavailable in France at the moment): Imagine With Meta AI. The company says that when photorealistic images are generated, the tool makes several arrangements. It places visible markers on images, but also invisible ones, via a watermark and metadata system in the image files. Concerning the metadata, they can be modified a posteriori to deceive: this method is far from infallible.

This is why at the same time, Meta is developing AI-generated image detection tools. Work that goes hand in hand with preventing the removal of watermarks: this is the technology in development at its laboratory dedicated to AI, FAIR, which is called Stable Signature. Meta specifies that the latter “ integrates the watermarking mechanism directly into the image generation process for certain types of image generators, which could be useful for open source models so that watermarking cannot be disabled. »