Artificial intelligence MIT Technology Review
Adobe has announced a new tool to help creators watermark their artwork and opt out of having it used to train generative AI models. The web app, called Adobe Content Authenticity, allows artists to signal that they do not consent for their work to be used by AI models, which are generally trained on vast… Nightshade and Glaze, two tools that let users add an invisible poison attack to their images. One causes AI models to break when the protected content is scraped, and the other conceals someone’s artistic style from AI models. Adobe has also created a Chrome browser extension that allows users to check website content for existing credentials.
Users of Adobe Content Authenticity will be able to attach as much or as little information as they like to the content they upload. Because it’s relatively easy to accidentally strip a piece of content of its unique metadata while preparing it to be uploaded to a website, Adobe is using a combination of methods, including digital fingerprinting and invisible watermarking as well as the cryptographic metadata.
This means the content credentials will follow the image, audio, or video file across the web, so the data won’t be lost if it’s uploaded on different platforms. Even if someone takes a screenshot of a piece of content, Adobe claims, credentials can still be recovered.
However, the company acknowledges that the tool is far from infallible. “Anybody who tells you that their watermark is 100% defensible is lying,” says Ely Greenfield, Adobe’s CTO of digital media. “This is defending against accidental or unintentional stripping, as opposed to some nefarious actor.”
The company’s relationship with the artistic community is complicated. In February, Adobe updated its terms of service to give it access to users’ content “through both automated and manual methods,” and to say it uses techniques such as machine learning in order to improve its vaguely worded “services and software.” The update was met with a major backlash from artists who took it to mean the company planned to use their work to train Firefly. Adobe later clarified that the language referred to features not based on generative AI, including a Photoshop tool that removes objects from images.
While Adobe says that it doesn’t (and won’t) train its AI on user content, many artists have argued that the company doesn’t actually obtain consent or own the rights to individual contributors’ images, says Neil Turkewitz, an artists’ rights activist and former executive vice president of the Recording Industry Association of America.
“It wouldn’t take a huge shift for Adobe to actually become a truly ethical actor in this space and to demonstrate leadership,” he says. “But it’s great that companies are dealing with provenance and improving tools for metadata, which are all part of an ultimate solution for addressing these problems.”
Published: 2024-10-08T13:00:00
