[Feature Request] Navigating the AI controversies

Software darktable
1

This thread is not intended as a new thread for discussion on the use of AI, the models etc. There is another thread and I think enough has been said.

However from that thread I want to point out that are a couple concrete requests to add to in darktable.

@Pascal_Obry

  1. would it be possible to have a feature flag that disables the inclusion of any of the AI stuff. I know - it is opt-in currently - but this would remove that opt-in.

  2. when a new release in made I - that is 5.6 - would it be possible to have an explicit non-ai release. So that users have a choice at download.

  3. would it be possible in when downloading the SAM model to inform the user that there are controversies surrounding the model.

This probably will not stop all the criticisms on the ai decision but at least users now have complete informed consent.

BTW: these are 3 different requests.

2 Likes
2

As far as I know, 5.6 will be released without AI. Anyone who wants that feature will have to compile it themselves.

1 Like
3

Could we have conformation on that from one of developers? Because as far as I know it is merged in the master??

4

Even if the code is merged compilation of AI is conditional. You need to use cmake switch USE_AI to generate corresponding Makefiles that include AI.

1 Like
5

Just to be crystal clear:

The default build, thus the eventual 5.6 release will not have any of the AI functionality. Therefore users that want to make use of AI have to build that themselves.

1 Like
6

You know that the code is public. Go check it out in GitHub. Not only you need to compile, you then need to download the models you want to use. As far as I know, the models have proper licenses.

2 Likes
7

There are two features flags actually:

  1. Conditional build with USE_AI.
  2. Runtime config to enable AI, which is off by default. If you do not touch it, no AI features will be available even if they were build.

So, if you don’t want and don’t need new AI features, just do nothing. They will not show up without you knowing.

5 Likes
8

The Windows, macOS and Linux AppImage binaries built on GitHub have support for AI but not a single model and not even the AI feature activated. So without the user doing something the AI feature are not visible and not usable.

For the binaries created by packagers (Debian, CentOS, Ubuntu…) we have no way to know if the AI support will be activated or not. That’s OS packagers decision.

Again my position but I expect the support for AI to be activated by default. That is, the application has no AI by default but at least it let the choice to everyone to enable it if needed. That’s freedom for all.

12 Likes
9

@Pascal_Obry @anry in the other thread there was a suggestion that is sensible to me to add to the AI preferences a link to a webpage (or maybe a large tooltip) showing where the models sit on the ethicality/transparency spectrum. I think it would be a nice addition to help users make informed choices.

2 Likes
10

Maintenance nightmare. Both for the link (which can go stale faster than you’d believe) and for the page itself. And even if the link stays current, the linked page can change to something less useful.
I don’t think it’s a good idea to have links pointing outside darktable within the program. Even feeling the need for such a link tells me there are still large issues with AI for general use…

Then, for the user to make an informed choice ( if they want to) they would have to do a bit more than just look at a single valuation (for which you would have to make sure the criteria are transparent as well!). Because that valuation in itself may not be objective (not saying it isn’t, just that it may be hard to judge).

2 Likes
11

That has been discussed, maybe not a link but some stats about each AI model like training dataset, copyright…

@anry : Is that still in your radar? Do you want me to open an issue to discuss where to show those information?

2 Likes
12

There is a good approach to this in the Linux kernel. The method could be adopted in a similar way. A kernel marks itself as “tainted” when something occurs that might be relevant later when investigating problems.

Darktable with AI could also be marked as tainted and in an unsupported state, as it cannot be guaranteed to function correctly due to external functionality.

13

@Masterpiga @Pascal_Obry I suggested similar model card per model, accessible in AI preferences. It is doable and makes sense as for me. I will work on that.

7 Likes
14

“tainted” is pejorative in English. Not the right word.

1 Like
15

Good solution. Let me know if you need help with the text. Happy to contribute.

16

People can already add LUTs that are data, just like the models; broken LUTs and broken models could lead to broken output. We never had a problem with that.

5 Likes
17

If everything is put into perspective, it ultimately becomes impossible to find answers to specific questions. This tendency to relativize is also used extensively in the other thread. This dilutes the topic and twists it into something absurd.

1 Like
18

I don’t understand why you’d treat one external data set, installed by the user and used by fixed algorithms to process the data, differently from another. How about lensfun data, then?

Besides, why should we claim that darktable is in “unsupported” state, because someone compiled the code with a configuration (build) option that is also used in the official builds, and enabled an option in settings that the official source code provides?

1 Like
19

AI is one of the biggest changes to our world for many years. It’s effect will be felt everywhere.
I’m not a coder or software developer and this AI constant discussion/argument does my head in.
It comes across as a knee jerk reaction.
I’m sure the many thousands of DT users will be bemused at these discussions. They probably get their updates on DT from Reddit rather than here.
Should a simple software package have such a moral stance?
If so keep it simple. At the release of future DT just give the option: with or without AI when downloading.
If I was a betting man I think I know what the options takes up would be.
DT is known for being a very active in it’s development and progress. It’s difficult to see how that can continue without AI influencing modules.

20

The most alarming things I have discovered about darktable from this huge AI discussion is that code contributors are not even asked to sign royalty-free licenses and that security code checks are voluntary at best. These are truly worst practices, if confirmed.