Meta has announced plans to label artificial intelligence (AI)-generated images on all its platforms, which include Facebook, Threads, and Instagram. This decision was unveiled on February 6, shortly after guidance from the company’s oversight board emphasized the necessity for an updated policy regarding AI-generated content. The call for action came in response to the dissemination of a digitally altered video featuring U.S. President Joe Biden. Meta clarified that while it currently labels photorealistic images produced by its AI technology, it will now collaborate with various companies to ensure that all AI-generated images shared on its services are properly identified.
In a statement published in the newsroom Tuesday, Nick Clegg, Meta’s President of Global Affairs, stressed the importance of marking AI-generated content to safeguard users and combat misinformation. He indicated that the company has initiated partnerships with industry stakeholders to establish a standardized method for signaling when content is AI-generated. “We’ve been working with industry partners to align on common technical standards that signal when a piece of content has been created using AI,” he remarked. The company also disclosed that it can currently label images produced by Google, OpenAI, Microsoft, Adobe, Midjourney, and Shutterstock, as well as those created by its own AI as “Imagined with AI.”
For accurate identification of AI-generated images, detection tools need a universal identifier embedded in such images. To this end, several companies in the AI sector have started employing invisible watermarks and embedding identifiers in the metadata of images, indicating they were not created by humans. Meta stated it can successfully recognize AI images from the aforementioned companies due to adherence to industry-standard technical guidelines.
However, challenges persist. Firstly, not all AI image generators utilize watermarking or similar identification techniques. Additionally, Meta has observed that various methods can remove these invisible watermarks. To address this, the company is collaborating with industry partners to create a standardized watermarking technology that cannot be easily stripped. Last year, Meta’s Fundamental AI Research (FAIR) team unveiled its efforts to develop a watermarking system named Stable Signature, which integrates the marker directly into the image generation process. Similarly, Google’s DeepMind has introduced a comparable tool known as SynthID.
The focus on image labeling is solely part of the broader challenge, as AI-generated audio and video have also become increasingly prevalent. In recognition of this, Meta admitted that a reliable detection technology for audio and video has yet to be developed, although progress is underway. Until an automated solution for identifying such content is available, the platform has introduced a feature that encourages users to disclose when they are sharing AI-generated audio or video. Once users provide this disclosure, the system will automatically add a label to the content.
Clegg further cautioned that failure to disclose such content could lead to penalties for users if Meta later identifies the material as digitally altered or AI-created. Moreover, if the shared content holds significant potential to mislead the public on critical matters, a more prominent label may be applied to assist users in understanding the context.
