Adobe Proposes 'robots.txt' for AI Image Training

Adobe's Initiative for Image Control in the Age of AI

Traditionally, website owners have utilized the robots.txt file to specify which web crawlers are prohibited from accessing their sites. Now, Adobe is introducing a comparable system for images, integrating a new tool into its content credentials framework. This aims to grant creators greater oversight regarding the utilization of their work in the training of artificial intelligence models.

The Challenge of AI Crawler Compliance

A significant hurdle may lie in persuading AI companies to genuinely respect Adobe’s proposed standard. It's worth noting that AI crawlers have a history of disregarding directives outlined in robots.txt files.

Understanding Content Credentials

Content credentials represent information embedded within a media file’s metadata, serving to verify its authenticity and establish ownership. This approach embodies an implementation of the Coalition for Content Provenance and Authenticity (C2PA), a widely recognized standard for content authenticity.

Introducing the Adobe Content Authenticity App

Adobe is launching a new web application, the Adobe Content Authenticity App, enabling creators to attach credentials to image files, irrespective of whether those files were created or edited using Adobe software. This app also provides a mechanism for creators to communicate to AI companies their preference against using specific images for model training.

The app allows users to associate their credentials – including their name and social media profiles – with a file. Up to 50 JPG or PNG files can have credentials applied in a single operation.

Leveraging LinkedIn Verification

Adobe is collaborating with LinkedIn to integrate with the platform’s verification program. This partnership facilitates the confirmation that the individual attaching credentials to an image possesses a verified LinkedIn profile.

While integration with Instagram and X (formerly Twitter) profiles is also available, these platforms currently lack a verification system for credential confirmation.

Signaling Preferences for AI Training

Within the app, users can select an option to indicate that their images should not be utilized for training AI models. This preference is then recorded in the image’s metadata alongside the content credentials.

Despite the inclusion of this field, Adobe has not yet secured agreements with AI model developers to adopt this standard. The company is actively engaged in discussions with leading AI firms to encourage adoption and respect for the indicator.

The Importance of Industry Adoption

Adobe’s efforts are well-intentioned in providing a signal to AI model developers regarding training data. However, the success of this initiative hinges on the willingness of AI companies to acknowledge and adhere to the standard.

Lessons from Meta's Experience

Last year, Meta’s implementation of labels to identify AI-generated images sparked controversy. Photographers voiced concerns about edited images being incorrectly tagged as “Made with AI,” prompting Meta to revise the label to “AI info.”

This situation underscored the differences in implementation, even among organizations like Meta and Adobe that are both members of the C2PA steering committee.

Creator-Focused Design

According to Andy Parsons, Senior Director of the Content Authenticity Initiative at Adobe, the new content credential app was developed with input from creators. Recognizing the fragmented regulatory landscape surrounding copyright and AI training data, Adobe aims to empower creators to express their intentions regarding AI platforms.

“Content creators are seeking a straightforward method to indicate their objection to their content being used for GenAI training. We’ve received feedback from both individual creators and agencies expressing a desire for increased control over their work in relation to AI training,” Parsons explained to TechCrunch.

Expanding Tool Availability

Adobe is also releasing a Chrome extension that allows users to identify images containing content credentials.

The company details that the content credentials app employs a combination of digital fingerprinting, open-source watermarking, and cryptographic metadata to embed metadata within an image’s pixels. This ensures that the metadata remains intact even if the image is altered. Users can then utilize the Chrome extension to verify content credentials on platforms like Instagram that do not natively support the standard, indicated by a small “CR” symbol.

Ownership and Attribution in the AI Era

In the ongoing debate surrounding AI and art, Parsons emphasizes that C2PA does not seek to define or dictate what constitutes art. However, he believes that content credentials can serve as a valuable marker of ownership.

“There exists a grey area – images edited with AI but not entirely AI-generated – and our goal is to enable artists and creators to sign their work and claim attribution. This doesn’t necessarily validate intellectual property or copyright, but it does signify that someone created it,” Parsons stated.

Future Expansion

Adobe has indicated that while the initial tool is focused on images, plans are underway to extend support to video and audio formats in the future.