Axios Pro Exclusive Content

Exclusive: Deep dive on AI and watermarking

Jan 4, 2024
Logos for IBM, and MetaAI are shown

Photo illustration: Jonathan Raa/NurPhoto via Getty Images

Watermarking is being tossed around by policymakers as a key way to help people identify AI-generated content, but a new paper says lawmakers should consider more authentication techniques as they develop legislation.

Driving the news: The paper from the Information Technology Industry Council shared exclusively with Axios argues that a combination of methods will be most helpful for consumers to navigate AI content, and that clear standards for authentication are needed for the industry.

  • Watermarking — embedding a "signal" in a text or an image to identify it as AI-generated — is just one way to tackle the problem, and the paper argues there are other useful ways to show a piece of content's roots that need to be part of the conversation as lawmakers develop new rules.

Other methods include:

  • Provenance, which traces "signals" in a dataset, such as when it was created and modified.
  • Metadata auditing, or checking things like editing history, time stamps and device information, which could be useful for copyright concerns, per the paper.
  • Human authentication, which would have experts decide whether something has been AI-generated. It would be slower and less reliable but useful in certain cases.

Yes, but: In some cases, marking whether something has been AI-generated may not be necessary at all.

What they're saying: "There's a strong need to examine which use cases and instances where watermarking makes sense, because lower-risk applications might not need that sort of disclosure, such as changing the filter or lighting mode on a smartphone's camera," Courtney Lang, ITI's vice president of policy, trust and data, told Axios.

What's next: Lang said it's too early to tell what type of content authentication method would be best for educating consumers about AI-generated content in elections.

Go deeper