Meta will begin annotating AI-generated photos uploaded to Facebook, Instagram and Threads in the coming months as elections around the world come together. The company will also begin penalizing users who don’t disclose whether real-life videos or audio were produced by artificial intelligence.
Nick Clegg, Meta's president of global affairs, said in an interview that these measures are intended to "incentivize" the technology industry as artificial intelligence-generated media becomes increasingly difficult to distinguish between true and false. The White House has been pushing hard for companies to watermark AI-generated content. Meanwhile, Clegg said, Meta is developing tools to detect synthetic media even if its metadata has been doctored to obscure the role of artificial intelligence in its creation.
Meta already puts the "Imagined with AI" watermark on images generated using its own ImagineAI generator, and the company will begin applying the same watermark to AI photos generated using tools from Google, OpenAI, Microsoft, Adobe, Midjourney and Shutterstock. Clegg said the industry was far behind other industries in establishing standards for identifying AI-generated video and audio. While Meta remains highly vigilant about how this type of media can be used to deceive, the company cannot catch everything on its own.
He said: "We will remain highly vigilant to those who are concerned that video, audio content is designed to deceive the public on important political issues leading up to the election. Do I think it is possible that no matter how quickly discovered or quickly labeled, we will still be inexplicably accused of being off the chain? Yes, I think it is possible, if not very likely."
Meta has been working with organizations like the Partnership for Artificial Intelligence to build on existing content authenticity initiatives. Adobe recently released a content credentials system that incorporates content provenance information into image metadata. After releasing a beta version of SynthID watermarking for images, Google has expanded it to audio files.
Clegg said Meta will soon start asking users to disclose posts with photorealistic video or audio produced using artificial intelligence. If users don't disclose, he said, they "will be subject to penalties ranging from warnings to deletion of offending posts."
There are already many examples of AI-generated viral posts from politicians, but Clegg underestimated the likelihood of the phenomenon taking over Meta’s platform in an election year. "I don't think we're going to see a completely synthetic, politically significant video or audio anytime soon," he said. "I just don't think it's going to be that way."
Meta has also begun testing internally the use of large language models (LLMs) trained to community standards, he said, describing it as providing an efficient "triage mechanism" for tens of thousands of human reviewers. "This seems to be a very effective and fairly accurate way to ensure that the content being reported to our human reviewers is indeed the kind of edge case that requires human judgment," he said.