Standardizing AI Iconography
We’re releasing a set of example icons to mark AI-generated content, in order to demonstrate a vision for an industry-wide effort to clearly label synthetic content.
Deepfakes and synthetic media are getting better, pointing toward a future where AI-generated or AI-manipulated content may be truly indistinguishable from real images, video, and audio. High-profile instances of misuse are already occurring, and it is clear that synthetic content will be used in grossly misleading ways.
We must promote a norm of disclosure:
Powerful tools and protocols like C2PA already exist which can use cryptographic signatures from cameras to validate genuine content. Meanwhile, watermarking can help platforms detect content which was produced by cooperating AI products. However, they need to be integrated with an informative visual standard in order to clearly communicate whether generative AI tools were used to create the content
We’ve provided an example set of icons for labeling both real and synthetic content to demonstrate our vision for a universal standard of disclosure. We hope this demonstrates the possibility for an informative, simple, and cross-platform visual standard for differentiating synthetic from real content.
The purpose of this visual standard is not to resist removal, but to suggest standard iconography that platforms and communities can use to disclose synthetic content to users. The watermarks for images, video, and audio that we expect will actually be robust to removal are not perceptible to humans.
Instead, we envision native adoption of this standard or a standard like this by platforms and communities, which already work to detect and moderate synthetic content. When a piece of content is flagged as synthetic, our icon can be natively tagged to it (for instance, as a moderator-only post flair on Reddit or added to the description of a YouTube video), where it cannot be removed without deleting the content. The reliability of a disclosure thus depends on the reliability of the underlying moderation or detection systems.