Skip to main content
A
angrymongoose
Known Participant
August 4, 2025
Open for Voting

P: Marking your photos as not to be used for AI training: c2pa "Do Not Train" assertions

  • August 4, 2025
  • 2 replies
  • 255 views

With the addition of Content Credentials, Adobe has taken a first step toward allowing creators to make use of Content Authenticity Initiative features. I'd like to suggest a second.

 

While researching Content Credentials, I found that the CAI defines a fairly extensive set of assertions that can be added to images (see https://opensource.contentauthenticity.org/docs/manifest/assertions-actions/). One potentially interesting set are the "Do Not Train" assertions, such as c2pa.ai_training, c2pa.ai_generative_training and so on, which allow creators to specify whether they want their images to be used for various kinds of AI training and data mining or not.

 

Because of growing concerns about images being used to train generative AI, a lot of creators would probably welcome the ability to express their preferences and 'warn off' AI scrapers from misusing their work. Much as Lightroom allows us to attach copyright information to images (via IPTC), it would be good if we could also add c2pa.mining_training assertions.

I see this as something that probably belongs in the export dialog, but it could also be implemented on a per-image basis. 

 

 

 

    2 replies

    Ｄ
    ＤＡＶＩＤＤＥＳＩＧＮ1988
    Participant
    October 22, 2025

    Subject: C2PA “Do Not Train” vs. Data Mining – Are These Technically Enforced?

    Hi Adobe team,😄 im DAVIDDESIGN-

     

    I’ve been researching the Do Not Train assertions inside the C2PA/CAI framework (like c2pa.ai_training and c2pa.ai_generative_training).
    From what I understand, these seem to act as declarative preferences rather than enforceable restrictions.

    However, since many AI datasets are built through data mining or feature extraction processes — not explicit “training” —


    I’d like to ask:

    1️⃣ Are these C2PA “Do Not Train” assertions technically enforceable at any level of data ingestion?

     

    2️⃣ Does Adobe plan to extend the standard to cover data mining / embedding / feature analysis use cases (e.g. c2pa.mining_training)?

     

    3️⃣ If a third-party crawler strips metadata before indexing, does Adobe still treat the original C2PA declaration as legally binding in provenance verification?

     

    These seem important questions, since without a cross-layer enforcement mechanism (between metadata and dataset parsing),


    the “Do Not Train” flag may remain symbolic rather than functional.

     

    Thanks for clarifying — I believe many creators would appreciate a technically transparent answer here...

    Terms & ConditionsAccessibility statement
    Using the community Terms of use Privacy Cookie preferences Do not sell or share my personal information ad choicesAdChoices