Digital Watermarking

Updated November 17, 2024

 

Digital watermarking uses technical methods to identify AI-generated content. For example:

  • The European Parliament published a briefing on generative AI, watermarking technology and regulation in December 2023.
  • Google DeepMind announced a watermarking tool SynthID for AI-generated images in August 2023, and extended and open-sourced it in October 2024.
  • Meta
    • in February 2024 announced that it is developing "common technical standards" (including C2PA and IPTC) for labeling AI-generated images on Facebook, Instagram and Threads
    • in July 2024 released its AudioSeal technology for watermarking AI-generated speech
    • in November 2024 released on Github "Watermark Anything" code for "embedding (possibly multiple) localized watermarks into images".
  • OpenAI
    • in February 2024 announced that it will use C2PA to identify images generated using ChatGPT or DALL-E 3

    • as of August 2024 reportedly had had for a year an effective tool for watermarketing ChatGPT outputs, but was unwilling to release it because many users prefer outputs that are un-watermarked (and therefore not attributable to AI).

  • California passed  legislation in September 2024 mandating that generative AI systems disclose that their outputs are AI-generated.