Punished for what? I totally understand that you might not like the direction I’m taking. You don’t have to install it everything is clearly explained in the README, and there are plenty of other forks available. I’m not going to get mad over that. The only thing I want is for the app to be on the Alt Store, even if it’s hidden.
I’d rather be honest about using AI than ship a slop app or vibe-coded mess while staying quiet about it.
My workflow is simple: AI is an assistant, even when it’s the one writing the code. I’m the owner of the project, so at the end of the day it’s my call whether something is good enough to merge or not. I’ve programmed before, so logic isn’t new to me, I try to keep things as clean and maintainable as I can.
To be clear on definitions: to me, a “vibe coded” app is one where the dev isn’t really the dev anymore, just prompts AI and ships whatever comes out, untested, usually with that same generic AI-slop UI. That’s not what I do. Try the app, test it, and judge for yourself.
I’m not losing sleep over this, and I get where people who are wary of AI are coming from. Either way, this project is made with passion, no ulterior motive, it’s a hobby.
People can do what they want with that info. My only goal here is to be on the store so other people can easily install my app, that’s it. If someone doesn’t want my app because of this, that’s fine too, they just don’t install it.
On a side note, what’s your actual take on AI in this kind of project? Genuinely curious to hear your view.
I respect the transparency. I was hoping for a similar level of transparency from the F-Droid team.
Whether it’s the product of prompting a chatbot, or an experienced developer using code auto-generation as an “assistant”, code written by a Trained MOLE is something many people who care about software freedom want to avoid. Whether its because of copyright licensing issues - the fact that MOLEs can spit up verbatim chunks of code they’ve ingested, and/or the murkiness of who owns the copyright on MOLE vomit - or because of social/ environmental concerns about the practices of MOLE Training in general. For the sake of transparency, I think the presence of auto-generated code ought to be declared by the F-Droid app, and we ought to have the option to filter out apps whose developers use this practice.
But this would need to be a separate filter from the existing one applying to apps that are tied to nonfree network services (eg using a Trained MOLE remotely). Also from the not-yet-existing one I think ought to apply to apps that contain a Trained MOLE as a binary blob (eg Fennec, apparently). It seems to me that this condition;
“Apps on F-Droid are free and open source (FOSS), they respect your privacy and they are available for anyone to see the source code, build on and improve.”
… is not met by apps that contain binary blobs the F-Droid team has not independently reproduced. In which case these MOLE blobs and the features that depend on them ought to be removed from F-Droid versions. Creating an antifeature flag so we can at least filter them out seem like a very reasonable compromise. To quote a recent comment on Lobste.rs;
"Open source ‘AI’ doesn’t exist. Locally-runnable LLMs are opaque blobs provided through the largesse of megacorporations who spend millions of dollars to train each revision and exercise total control over what goes in. As an individual, non-independently-wealthy person you cannot inspect everything about how they’re constructed, tinker with their original training set, and rebuild them from scratch at will. Running a necessarily pre-compiled gratis LLM on your personal computer is an act of cultivated dependence upon centralized infrastructure and an endorsement of technology which- in our present, real world- structurally cannot be controlled by individuals.
I’m finding it very concerning that no one from the F-Droid team seems to be willing to seriously engage with the implications of shipping these proprietary components, in an app library that exists to filter out proprietary software and the various abuses it enables.
I agree @vdbhb59 about this. The question asked was whether the F-Droid project has an explicit policy on the “AI” issues, and implicitly, whether they’ve analysed these issues in any detail. It’s pretty clear that the answer to both questions is “no”. I can see value in opening new threads that ask more specific questions, or make more specific proposals, informed by the discussion here so far. The mod team may disagree and merge them back into this one 
Just letting people know of this blog post posted today: How much of F-Droid is LLM generated? | tintotint's site
It only looks at one batch of updates, 102 apps (so don’t think that the apps not listed are not vibe-coded, I know myself of many that aren’t listed there).
Read through it, thanks for putting the time into this. For what it’s worth, N-Zik got the “Mostly AI” tag there, which tracks with the disclosure I’ve already got in the README (AGENTS.md, etc.). I don’t think that’s an unfair read given the criteria used (agentic infra = auto-tagged).
That said, it’s worth noting the piece is explicitly a “rough 3 tier system based on the aesthetics of the repo” rather than actual code analysis the author says as much, and that’s fine for a blog post, but it’s a different bar than what an actual F-Droid AF/tag policy would presumably need (verifying human review happened vs. just looking at commit co-authorship). Might be useful data for this thread regardless, since it gives a rough sense of scale (72.5% “mostly AI” out of that batch) rather than just anecdotes.
As a user, I am interested in knowing two things:
- Is the app written in significant part or in full using generative AI?
- Does the app employ AI to do its function? e.g. Chompass
- There is a third one, “Is the app about AI?” but it is more of an issue on desktops currently. See new FOSS apps on alternativeto.
Point 1 is about an ethical stance for some people, but it is also about code quality. It is not a sure sign but the association is high.
Point 2 touches on privacy and usability. Do I need to connect to a provider? Do I have to run a model on my RAM-limited phone?
I think these two are worthy tags to have disclosed by the developers. However, the term anti-feature is a bit loaded since some other people have no problem with either point. In any case, I think these tags, if added, should be unselected by default, or the user be offered a choice on F-Droid’s first install.
A good alternative UX may be to show a switch in the app page next to any “anti-feature” that says: “Don’t show any more apps with this tag”.
I’m reviewing this thread while looking for a link posted here, and I noticed a comment that I didn’t catch on the first read;
The suggestion seems to be that a full generative model is equivalent to an image, or video, or font. These are the sorts of things covered by the NonFreeAsset flag. I don’t think this holds water.
Let’s say I create a proprietary multi-function software engine. I compile it into a blob and have a copy downloaded and included in my otherwise Free Code app, either when its compiled, or when it’s installed. Would this app be suitable for inclusion in F-Droid? Because that’s effectively what we’re talking about here.
Not sure how they are related. Model is not software engine.
omg you guys are gonna actually end up with one of these policies that just AI marketing and do nothing about the binary blobifant in the room, can’t wait
In my opinion:
-
Vibe coding should be forbidden in any open-source project, especially for F-Droid.
-
Proprietary AI should not be used. Only open-source AI should be allowed for checking bugs or similar tasks, but not for coding.
-
If OpenAI is used for checking bugs or similar tasks, this should be disclosed on F-Droid.
I am not sure that I understand the principle here. To me it reads as though no proprietary tools should be used to create software that is distributed by F-Droid. Does that include: compilers, IDEs, hardware, operating systems (including Android itself)?
A generative model is a binary compiled from source code, which instructs the computer in performing a group of functions, as part of a larger program in which it’s embedded. A software engine is exactly what it is.
Please explain the logic of why why think is isn’t software.
The problem here is not the same as someone writing code on a MacBook, or in a nonfree IDE. They have never been known to insert copyrighted code (proprietary, Source Available, or copyleft licensed) into the software being worked on. Code auto-generators do. That’s one difference among many.
That AI code generation can insert copyrighted code does not necessarily mean that it has, in any given software project. It might happen, it might not. Just the same as using a any text editor (proprietary or not). The developer can insert copyrighted code (deliberately or unknowingly). And certainly a compiler (proprietary or otherwise) can do that (in the resulting executable code) without any knowledge of the developer.
It is the resulting code that matters, not the tools that were used to develop the code.
I still cannot see the principle in banning software solely owing to the developer having used AI somewhere in the development process. If it can be shown that the resulting code is violating copyright, then sure. Having used AI is not proof of that having happened.
LLM is not compiled from source code. It can only give an output from an input. It can’t instructs the computer.
wow what an amazing beast this LLM is… takes input and produces output… i wonder how it manages to do that… mystery of the ages… but wait a minute! could it be?! maybe pehaps it might be that it’s a set of instruction, like arithmetic operations maybe, that you computer can perform in order to transform the input into the output!!! noooo, it can’t be, what a silly idea, ha ha
also yeah all of you guys in this thread too should lookup “preferred form”
Woah. Thanks for sharing. The read was an eye opener for me. I had no idea that Feeder was would end up in the AI list, but I looked at the git repo and there it is, lo and behold, the agents.md file.
Not to say I’m disappointed, however, I’d prefer if the developers declared up front their usage of AI (if any) in their README.md. This would definitely help build more trust with users, given that there is an audience for no-AI. I wouldn’t mind if app dev is AI-assisted, especially if we’re dealing with a one-man band.