Flathub's AI slop ban looks like it was the right call

When Flathub banned AI-coded app submissions last month, some critics warned the platform was denying the future by dismissing a new wave of “vibe-coded” software as out-and-out “slop”.

Well, new data suggests otherwise, as nearly three-quarters of the rejected apps are already dead, existing only a few months.

Linux developer Evangelos Paterakis, developer of Tuba, Turntable and others, did the digging, looking at 120 code repositories whose pull requests for inclusion on Flathub were rejected because of their heavy AI usage and given an “AI Slop1” tag.

Of those rejected, 88 are no longer under active development, many having deleted their source code entirely. Just 32 still show signs of life, but it makes for a 73% abandonment rate overall.

It nixes the idea users, so keen to benefit from a new wave of AI-coded wares, would rush off to use other formats and app stores since, it turns out, whatever might have been swept up in the ban didn’t appear to be anything anyone felt bothered about maintaining anyway.

Hips don’t lie, but stats…?

The methodology used here was not rigorous or academic. Paterakis acknowledges this themselves in a blog post titled “Democratizing Abandonware“ (which is terrifically subtitled: “I’m not digging through this cemetery”).

And, sure, abandonment attrition isn’t unique to AI (though one could argue the one-shot nature of might amplify it) as open source has always been a bit of a graveyard for solo hobby projects, irrespective whether a chatbot helped to write them or not.

I know I’ve spent nearly 16 years writing about cool new software or distros which get people excited, only to succumb to cobwebs and crow beaks before the year is out.

But Paterakis findings are not apropos of nothing either.

They back up the real reason why Flathub felt a ban was needed: that volunteer reviewers would become swamped reading reams of code a submitter hadn’t, and giving feedback to machine agents happier to reply with confident word salad than actioning the fixes asked of it.

Slopcode and lazyware

Debate can rage on what constitutes software slop – for me, 20 music players that feel like increasingly garbled versions of Amberol, passed down through a game of telephone, kind of comes close – but I did see a musing from Thom Holwerda at OSNews I liked.

He said that if AI tools really have allowed developers to focus more on quality and innovation – as the marketing would have us believe – to build things that were previously out of reach, surely we’d expect to have seen some evidence of that by now.

So far, much of what AI has produced is merely a sandcastle simulacrum: simplistic, built quickly, then left for the tide to take. Pomodoro timers, note-taking tools, wrappers around local AI models – things that aren’t hard even for novice coders to create.

There’s scant sight of any really complex or truly unique/novel software washing up on Linux’s shores just yet. Perhaps. like the big returns on AI investment, those be along any day now, as soon as the next data center is online (or in orbit, or chained to a blackhole or whatever).

Vibe-coding has made it easier and more fun to turn an idea into an app, but not easier in having an idea or maintaining the enthusiasm to keep at it. AI is a tool, and like all tools, best used with grounded skill2 than lazy optimism.

Flathub volunteers add this tag manually to any pull request where the submitter appeared to be using chatbots to communicate with developers/respond to their feedback, or who submitted automated manifests, etc – tell-tale signs that things didn’t have a human in the real loop, basically. ↩︎

Nobody becomes a chef by microwaving a ready meal. But a chef knows exactly when a microwave is the right tool to help them make a great dish – same is true with AI. ↩︎