As it stands right now, users do not have to make a conscious choice to opt-in to include the history stack, geo-location data, exif, metadata, or tags. It doesn’t require a conscious decision to opt-in to include the fact that dt is the software used to edit. Dt makes a lot of choices (defaults) which people may explicitly not want, and it is up to the user to change. What matters is the why of the default.
Why is including geo-location not a conscious choice, but ai use is/should be? And how does that decision by the dev team interface with the areas of the photographic community which are working to find ways of keeping photographers from hiding ai use?
I think offering some sort of nod to that community, even if it is a mere fig-leaf, would be good. I get that others have very different opinions; and that may include a large percentage of the core devs.
Geo-location is part of the image metadata, if you have it enabled in the camera. At least in all of my cameras, it’s off by default and I haven’t enabled it.
If something were to be added to DT to automatically add an indicator in EXIF data that AI was used whenever certain modules (or module options if relevant) are used, then it would be important to identify, at a minimum in the doc but probably elsewhere as well, what modules/options will trigger the indicator. Otherwise, someone might submit an image to a contest that forbids AI use without knowing that they used something that made use of AI.
Regardless of anyone’s personal opinion of AI, it’s necessary to also think about the opinions/rules of that person’s audience.
I think the concern about AI transparency in photography is valid – particularly in photojournalism and competitions where authenticity matters. But I think the conversation only makes sense if we’re precise about what kind of AI we’re actually talking about.
There’s a meaningful difference between:
Generative AI – tools that synthesise new content from training data (generative fill, object removal, image upscaling via super-resolution). This is what most competition rules target, and rightly so.
Enhancement AI – tools that automate decisions a human would otherwise make manually (AI masking, neural denoise, subject detection). The output is fully determined by pixels the photographer captured.
Major competitions that have actually thought this through – World Press Photo, the World Photographic Cup – explicitly permit the second category while banning the first. And they verify compliance by requesting the original RAW file, not by checking metadata flags.
Which is why I’m sceptical that automatic metadata tagging would actually help:
Competitions that care already have a verification workflow (RAW file submission) that doesn’t depend on metadata at all.
Competitions that do blanket “no AI” bans are unlikely to have the tooling or expertise to interpret a tag like ai_masking_used: true correctly – it would look the same to them as generative fill, even though in effect it’s no different from drawing a mask by hand.
The metadata is trivially removable. Anyone trying to hide something will strip it in seconds. It only burdens honest users.
If darktable ever wants to engage with provenance meaningfully, C2PA seems like the right framework – a cryptographic, tamper-evident standard backed by Adobe, Canon, Nikon and others. It’s a significant undertaking, but it’s the only approach that can’t be circumvented trivially. The XMP sidecar already gives a complete auditable history of every module applied, which is more than most applications offer.
A photographer’s choices about In-camera capture data has nothing to do with dt’s default choice architecture. People may want that for personal use, but want it disabled for exports.
It is my impression (warning: small sample size lol) that this idea is starting to become more and more prevalent inside the photographic community as a whole. With notable exceptions of course!
re: C2PA
While C2PA is a worthwhile endeavor, I think it will run into uptake problems outside of a few very specific use cases.
If I understand it correctly (and maybe I dont!), this requires the camera to sign the RAW. The likelihood that manufacturers are going to roll out the necessary firmware to all previous models is (to me) virtually zero. I don’t know if the hardware in my Canon 50D is even capable. Film cameras will not be covered at all. So digitalization of the negative has to be done with a qualified camera unless all the scanner manufacturers decide to sign their outputs as well. I think this will be a gate that will limit adoption.
I can see some journalistic organizations and some competitions using the C2PA standards (if/when it gets off the ground), but the wider photographic community just does not have that kind of homogeneity. There isnt even homogeneity in competition rules or verification methodologies. Contrary to @anry 's assertion, some competitions use metadata even when they require RAWs at later points. Moreover, competitions are not the only places where there is an interest in regulating ai usage. Lots of people care about lots of differing things.
As has already been noted, C2PA is a big lift. Is that worthwhile? Sure! If it means photojournalists can use dt and be in compliance with their organization’s standards (where otherwise they could not) I would think that a good thing for the project. I don’t think it is even remotely able to address the needs of the wider photographic community, and multiple methodologies is the best way forward.
It doesn’t have to be metadata tags as in the OP, even if it is an easy (but wildly imperfect) solution, but something needs to be put in place. Or not…there is no shortage of sufficiently capable raw development platforms out there and people can always fork dt so, at the end of the day, this is a small ball set of problems. If dt gets in a photographer’s way they can just not use dt. It is the dev’s world… we just live in it.
I mostly agree, but with the exception of de-noise which I though was a generative process. I may be miss-understanding how it’s achieved in dt however.
I don’t think the C2PA metadata needs to be added at the camera, but it can be.
It’s a very useful tool to check the provenance of the image or video (when and where it was shot and by whom) and a list of what edits were made when. It also allows tagging of AI manipulation, the various software tools used and who has used the tools.
As has been said, the C2PA can be replaced - you could imagine a newsroom using it.to.check it’s provenance and then replacing it with their own data (there are valid reasons to do this, such as removing the location of a journalist who is embedded in the front line).
With careful design, the C2PA edit list could be a drop-in replacement for the sidecar file.
