May 6, 2024 - Technology

How Big Tech labels AI

Illustration of a framed painting with sticky notes of different shapes and sizes all over it

Illustration: Natalie Peeples/Axios

Tech platform policies on labeling AI-generated images and video users are a moving target.

Why it matters: The explosion of generative AI online is making it harder for people determine what is and isn't real on social media.

Between the lines: Tech giants have guidelines in place for their own platforms, and are also joining coalitions to set industry standards for labeling.

  • Meta, Google and TikTok all say that content made with their own AI tools will be labeled automatically.
  • But content that's made with other AI tools and posted on these platforms is harder to label.

YouTube currently requires disclosure for any "content that is meaningfully altered or synthetically generated when it seems realistic."

  • "Beauty filters" are OK, but generating a whole new face is not.
  • If users of generative AI don't label their own work, YouTube says it'll apply the labels for them and potentially remove the content or suspend the user from YouTube's partner program.

TikTok requires labeling for any content that "contains realistic images, audio, and video."

  • That means content that's been edited "beyond minor corrections or enhancements."
  • This includes showing a subject doing something they never actually did or saying something they never said. It also includes any use of face-swapping apps.

Meta released new guidelines earlier this month based on feedback from its Oversight Board.

  • Original guidelines restricted content that was manipulated to depict fake speech, and now also covers content altered to falsify someone's actions.
  • Meta says it will use "industry-shared signals of AI images," advice from fact checkers and self-disclosure when it starts labeling content in May.

Yes, but: It's unclear whether any of these companies have the moderation support to enforce these rules and standards.

  • A recent study from the Stanford Internet Observatory found that Facebook's algorithm was feeding AI-generated content (usually unlabeled) from accounts that people don't follow because the engagement on those posts was so high.
Go deeper