On February 6, Meta announced that she would label AI-generated images on Facebook, Instagram and Threads. When someone uses Meta's AI tools to create images, the company adds visible markers to the image, as well as invisible watermarks and metadata in the image file. The company says its standards align with best practices established by the Partnership on AI, an AI research nonprofit.
Big Tech is also backing a promising technical standard that could add a “nutrition label” to images, videos and audio. Called C2PA, it is an open source Internet protocol that relies on cryptography to encode details about the origin of a piece of content, or what technologists call “provenance” information. C2PA developers often compare the protocol to a nutrition label, but one that indicates where the content comes from and who – or what – created it. Learn more here.
On February 8, Google announcement it joins other tech giants such as Microsoft and Adobe on the C2PA steering committee and will include its SynthID watermark in all AI-generated images in its new Gemini tools. Meta says he also participates in C2PA. Having an industry-wide standard makes it easier for companies to detect AI-generated content, regardless of the system it was created with.
OpenAI also announced new content provenance metrics last week. It says it will add watermarks to the metadata of images generated with ChatGPT and DALL-E 3, its image-creation AI. OpenAI says it will now include a visible label in images to indicate that they were created with AI.
These methods are a promising start, but they are not foolproof. Watermarks in metadata are easy to bypass by taking a screenshot of images and simply using that, while visual labels can be cropped or removed. There may be more hope for invisible watermarks like Google's SynthID, which subtly changes pixels in an image so that computer programs can detect the watermark but the human eye cannot. It is more difficult to modify them. Additionally, there is no reliable way to label and detect AI-generated video, audio, or even text.
But it's still useful to create these provenance tools. As Henry Ajder, an expert in generative AI, told me a few weeks ago when I interviewed him about how to prevent deepfake pornography, the goal is to create a “perverse customer journey.” In other words, add barriers and friction to the deepfake pipeline to slow down the creation and sharing of harmful content as much as possible. A determined person will likely always be able to circumvent these protections, but every little bit helps.
There are also many non-technical solutions that tech companies could introduce to avoid issues like deepfake pornography. Major cloud service providers and app stores, such as Google, Amazon, Microsoft and Apple, could ban services that can be used to create non-consensual nude deepfakes. And watermarks should be included in all AI-generated content, even by small startups developing the technology.
What gives me hope is that alongside these voluntary measures, we are starting to see binding regulations appear, such as the EU AI Act and the Digital Services Act, which require tech companies to disclose AI-generated content and remove harmful content more quickly. There is also renewed interest among American lawmakers in adopting binding rules on deepfakes. And following President Biden's AI-generated robocalls telling voters not to vote, the U.S. Federal Communications Commission announcement last week, it banned the use of AI in these calls.