Identifying AI contributions to EPUB content

[reidmore-online]

Section

https://www.w3.org/TR/epub-34/#sec-pkg-metadata

Describe the problem

AI/LLM usage is becoming prevalent within the publishing industry, and there are a variety of use cases where AI is used or can be used in the publishing process. However, due to known concerns with AI content generation, users have voiced interest in knowing how AI is used in their content. Knowing AI was used in content generation can impact a user's buying decisions or ways they interact with content, so it is important we offer a means for them to discover this information.

We don't currently have a means in EPUB metadata to clearly identify AI contributors or AI generated content within the book. We also may want to consider whether identification is required within content files, such as an attribute or other means to identify a particular piece of content as AI-generated.

Describe the fix or new feature you propose

ONIX has documentation on the related codes for this:
https://www.editeur.org/files/ONIX%203/APPNOTE%20Aspects%20of%20AI%20in%20ONIX.pdf

ONIX has codes for several types of "unnamed person", which can be found in list 19: https://ns.editeur.org/onix/en/19

The solution or approach for this problem involves two areas, changes to metadata and additions to markup. In the metadata case, this is well within our scope. In the case of markup, we likely want to come up with a proposal or rough sketch of what we want, and then take it to groups like AGWG or WHATWG as this is not only a publishing use case. We need to start the discussion somewhere.

[sueneu]

This is a timely issue.

On the image side, I was reading up on ChatGPT and noticed they embed C2PA metadata in images created by their DALL·E 3 model

If we see that other AI models are adopting this, could RS pick up and display that data?

[llemeurfr]

I read the C2PA technical spec some time ago. Some will certainly disagree, but I found it quite complex. Few software libraries allow parsing it. Therefore, I don't think RS will ever be able to make use of it. In the same vein, RS don't extract EXIF, IPTC or XMP metadata from images.

However, extracting such information from images on the production side would be a good idea. An EPUB metadata could therefore expose to RSes the corresponding C2PA metadata. Lots of wrok has been done on C2PA for representing such information, it would be sad not to use it.

[llemeurfr]

And more than that, there is a question of responsability and time of "divulgation". It is publisher's responsability to provide this information, not the RS by inspection of assets; and the best time for providing this information is when the user wonders if they will buy/borrow the book. In this case, ONIX is the optimal vehicule.

[iherman]

This was discussed during the pmwg meeting on 27 November 2025.

View the transcript

Ways to identify inferred metadata - w3c/epub-specs/discussions/2830

<wendyreid> w3c/epub-specs#2833

gautierchomel: I have questions about ways to infer metadata
… people are starting to update their epubs and that comes with a lot of AI generated content

gautierchomel: if we cannot differentiate what has been auto-answered it will become problematic

Hadrien: I raised this during tpac - there could be legal reasons for identifying what is AI generated
… might not be in our scope to do this for html
… if there is work going on in this area it would be helpful to know where

ivan: w3c is still trying to find out what can be done about AI - there is a new interest group but not answer at this time
… even if we are talking metadata this is not just an epub issue
… coming up with our own metadata in isolation is not a good idea

wendyreid: we might be overthinking the question - this is only epub metadata and what is generated automatically

<Hadrien> https://w3c.github.io/pm-wg/minutes/2025-11-11-f2f.html#2ee0

wendyreid: how can you tell if a book contains AI-generated metadata
… could we use something like the refines attribute to make statements about resources
… all we are doing is putting an asterisk on resouce essentially

Hadrien: I think ivan's point still stands that we don't have the metadata to say anything - we're lacking the vocabulary and it's not in our scope to define that
… could be schema.org or dublin core or someone else to define the metadata

wendyreid: if I have a million epubs in my database and some are quite old and I want to add accessibility metadata then I can try to infer things like whether they support some textual reading
… but there should be a way to say that this metadata was automated so it is not compeletely trustworthy

gautierchomel: the problem is bigger, but yes, that is what we need to capture first

---

[gautierchomel]

This is also to be considered as part of the discussion about #2830

i.e. we need to be able to identify what was automatically added, and when it is the case if that automation implies AI.