Re: Minutes of the W3C CG meeting of 25th Jan

Leonard - in your answer, you didn't respond to Giulia's question:

> Furthermore, it seems that the V2.0 of this specification deprecates the "Training and Data Mining" properties previously defined in the V.1.3.  Could you please point us at the new way of expressing tdm opt-out in C2PA? 

Having the opt-out section deeply buried in the current version of the C2PA spec, or worst, absent from this version, could let people think that C2PA is not interested in opt-out after all.

Laurent


> Le 26 févr. 2024 à 15:09, Leonard Rosenthol <lrosenth@adobe.com> a écrit :
> 
> Giulia – as an international organization, the W3C should be focused on developing and fostering standards that support the needs of ALL countries around the world, not only those of a single one.  While we could certainly debate which solution better aligns with the EU DSM Directives for Text and Data Mining, many other countries have other requirements that need to be considered as well. That may be why the US, UK, AU, China and others have focused on C2PA as the technology of choice for identification of material sourced from a generative AI system, and specification of whether one can use that asset for training.
>  
> In addition, as mentioned the PDF industry (as driven by the PDF Association – pdfa.org <http://pdfa.org/> – and ISO TC 171/SC 2) has already chosen to align with C2PA and implementations from Adobe and others are already in the field.  In addition, since the same solution/technology is also in use by numerous systems around the world for images and videos, it makes sense to land on a single choice for all media types.
>  
> There is no question that the work of this CG was instrumental in moving the industry to understand the problem space and the importance of addressing it – in fact, it certainly helped me to design the solution present in the C2PA specification.  However, the actual technical solution hasn’t received the industry support while others have.  It happens in the world of standards…
>  
> Leonard
>  
> From: Giulia Marangoni <giulia.marangoni@aie.it <mailto:giulia.marangoni@aie.it>>
> Date: Thursday, February 22, 2024 at 11:42 AM
> To: Leonard Rosenthol <lrosenth@adobe.com <mailto:lrosenth@adobe.com>>, public-tdmrep@w3.org <mailto:public-tdmrep@w3.org> <public-tdmrep@w3.org <mailto:public-tdmrep@w3.org>>
> Subject: R: Minutes of the W3C CG meeting of 25th Jan
> 
> EXTERNAL: Use caution when clicking on links or opening attachments.
> 
>  
> 
> Dear Leonard, 
>  
> We are aware of the C2PA specifications. Unfortunately, as discussed in our previous exchanges, we do not believe they currently meet some of the core requirements that underpins the TDMRep initiative, such as compliance with the conditions of Article 4 of the EU DSM Directive on Text and Data Mining. 
>  
> As you know, the DSM Directive introduces an exception for text and data mining which is relevant in the context of AI. Consistently, as confirmed by the EU Commission, text and data mining processes are understood to cover AI training and the TDMRep Protocol is following this assumption to ensure legal compliance. This approach has recently been confirmed by the EU AI Act, which requires AI providers to actively identify and respect TDM opt-outs, as defined by the DSM Directive, when training an AI system. It also clarifies that any AI made available on the EU market would have to comply with its rules, including on copyright, regardless of where the AI was trained (making opt out rules not only relevant in Europe but beyond).
>  
> Besides, it is worth considering that the TDMRep is intended to be a "format agnostic" solution, meaning that the same mechanism will be applied to different content formats, such as web, EPUB and PDF. This makes the TDMRep a horizontal solution across different formats - including but not limited to PDFs - developed in response to the needs expressed by the wider publishing sector, and not only by STM. Similarly, the TDMRep could be adapted to further formats, including beyond text, should the relevant industries express such needs.
>  
> As anticipated in the minutes, the draft specifications for embedding TDMRep in PDF are in progress and will be circulated and discussed at the next meeting of the W3C Community group as soon as ready. As far as we currently understand, the specifications to embed TDMRep in PDF do not create a technical issue that would prevent a PDF file from working as intended or that would make the opt out impossible to detect. We will of course welcome your views and demonstration, once the specifications are ready and circulated, should you believe that it would “break” a PDF file.
>  
> Furthermore, it seems that the V2.0 of this specification deprecates the "Training and Data Mining" properties previously defined in the V.1.3.  Could you please point us at the new way of expressing tdm opt-out in C2PA? 
>  
> Hope this clarifies, 
>  
> Giulia
>