Tag for applications that have used LLMs in development

[danirabbit]

The same way we have the ability to make choices about licenses that we're okay accepting, I think we need a way to mark apps which have been developed using LLMs like Claude Code so that folks can have an informed choice and avoid using this software on this basis if desired.

Unlike #652 which seems to be asking for a way to tag apps which interact with LLMs as part of how they function, this is about the app was made and how its community conducts itself. So I don't think this can be covered by the OARS case as pointed out in that issue

[GeopJr]

I've also been thinking of this and ended up with two different approaches:

1. Freeform AI disclosure
2. OARS-like categorization: "code", "translations", "branding"...

The first one gives developers the ability to fully describe their use of LLMs, for example "Used Clause to generate the CI config". While the second one allows stores to display them like the permission dialogs and summarize how harmful they are (again, like with permissions).

My personal take is that some uses like in translations are not equivalent to using Claude to generate an entire app and a general "Developed with LLMs" badge wouldn't do it justice.

[danirabbit]

That's a great point. I'm also thinking of the kind of "nutrition label" pattern like in the iOS app store where you can get a pretty detailed list of exactly how data is used, something like this for LLM, Generative AI, and ML use could be good:

https://developer.apple.com/app-store/app-privacy-details/

I agree there's a severity difference between something who uses Claude Code vs someone who generated their icon vs someone who used machine translation. I think an OARS style list of tags is a great idea

[danirabbit]

In terms of prior art I found https://daiu.org/ which is a simple AI disclosure standard: https://responsible-ai-usage.github.io/daiu-spec-site/

They’re mainly focused on trying to give a quick Creative Commons style tag which asks two questions:

• Origin: [Human created, AI Generated]
• Involvement: [None, Assist, Remix, Major, Full]

[gvisoc]

At this point I agree that there must be a disclosure, and I also have a lot of questions that grow by the minute. Apologies for the length.

1. What about local LLMs? They are useful and sometimes not that notoriously disclosed as the Claudes and the GPTs. Depending on the tone of the resulting position some may think that this is biased just against cloud models, which I don't think it'd be the intent.
2. Should there be something about contributing / including the context files as part of the source code? Happily prompting to produce apps end to end produces very low quality slop, while good context files often doubles as good documentation that can be used to improve the (assisted) code or allow contributors to leave LLMs behind. The same that free software includes the source code, assisted source code should include the context used for the models that contributed or helped the author. Otherwise it's a step back into obscurity and feels closer to closed source.
3. Same goes for the rules / configuration of the models or agents. They are good starting points for a contributing.md.

On the last 2 points, I think that if a developer that uses LLM (local or not) for assistance and does not publish the agent rules and the context files, but only the generated code, is not acting in good faith.

[UncleTallest]

I like the OARS-like categorization idea... I love the daiu spec.

@gvisoc context can contain PII so that's kinda dangerous. that would have to thought through carefully.

[gvisoc]

I've never seen a context file with PII, but it definitely should be reviewed. Mixing context oriented to generate code with secrets or any other form of sensitive / PII data is sort of the same mistake as committing your API Keys.

[danirabbit]

It might be worth specifying if local or cloud models were used since the data centers used for cloud models are a primary ethical concern. But then again even if you run the model locally the training data was almost certainly not obtained locally so I’m not super sure that matters either. This also kind of seems like getting into expecting regular people to make engineering decisions which is not really where I want to see this go.

I’m not really sure how much it matters if you check in the prompts/config/context or not since it is not a compiler and the output is non-deterministic (even though it seems like the goal of those files is to try to get something more deterministic). When you check in source code, you can compile it and you will get the same output as the author since the compiler is deterministic. I don’t think you can license or copyright a prompt either. So it seems not really relevant to me what the prompt was if the source is available 🤷🏻‍♀️

I think we need to be able to really concisely explain why something matters in a way a nontechnical person can understand and if you can’t, then that probably isn’t relevant

[UncleTallest]

Context is chain loaded from start prompt to the actual prompt used for a given task. Most of the first 5 or 6 items in that chain are designed to tell the model instance who the user is and anything about them the instance needs to know. That's all context. I agree about potentially seeing the actual prompt used to build an item; but that prompt is only the tip of the context iceberg. Just my two cents... From: Gabriel Viso Carrera ***@***.***> To: ***@***.***> Cc: "Uncle ***@***.***>, ***@***.***> Date: Mon, 23 Feb 2026 20:22:48 -0600 Subject: Re: [ximion/appstream] Tag for applications that have used LLMs in development (Issue #744) gvisoc left a comment #744 (comment) I've never seen a context file with PII, but it definitely should be reviewed. — Reply to this email directly, #744 (comment) , or https://github.com/notifications/unsubscribe-auth/AAEYP77FCVI7H6TGIY3WKAD4NOYXRAVCNFSM6AAAAACV5KDMPCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTSNBYGUYTCMZUHA . You are receiving this because you commented. The content of this email is confidential and intended for the recipient(s) specified in this message only. It is strictly forbidden to share any part of this message with any third party, without the written consent of the sender. If you received this message by mistake, please reply to this message and follow this with its deletion, so that we can ensure such a mistake does not occur in the future.