As of writing, I have the particular privilege of working on open source code as a full-time, salary-paying job.
I have to attach a conditional to that because the open source nature of it isn’t a guarantee—what I do could get put behind a wall next week for all I know—it’s only open because someone higher in the organisation allows it to be.
Being able to do stuff in the open and witness how it spreads, mutates and feeds back on itself is a major draw of the job for me. Whilst I’m no Linux-headed evangelist, I do love open source software.
I also sincerely think that generative AI (genAI) is an existential risk to all open source software projects.
AI is the thief of joy
One of the supposed ‘selling points’ of genAI in development is that it makes code creation easy.
Rather than writing code yourself, you can write a detailed prompt and feed it into a machine that will magic the code out of thin air instead, saving effort (if not money) and leaving more time to write new specifications and review the generated outputs.
If you’re any kind of impassioned developer, this won’t be a selling point to you because writing code and solving practical issues is the fun part of the work. Writing specifications and performing code reviews are the necessary evils of writing production software; not the bits that developers actually want to spend most of their time doing.
This is especially relevant in open source, where most people are only contributing because contributing is something they enjoy doing. This year’s State of the Devs survey reflects that, with a plurality of respondents saying that AI has had a negative effect on their mental wellbeing, and a majority of those saying it has made them exhausted and disillusioned with the industry.
If you take away the fun aspects of developing software, no one will want to develop software anymore. It’s that simple.
AI isn’t even good at writing code
Now, I’ve used generative AI before, I don’t go in for criticising a technology without having tried it, after all. I’ve tried the autocompletes, agents and assistants and generally didn’t find them useful, or not useful enough to justify the myriad downsides they come with. (I’ve written about this before.)
I’m a frontend developer. My day job is writing HTML, CSS, and JavaScript (or superlanguages therein) with a particular orientation towards accessibility, progressive enhancement, and semantics.
No code-writing AI I’ve tried has ever output frontend code that I consider to be of acceptable quality. Never.
This year’s WebAIM report on web accessibility directly attributes a sharp uptick in inaccessible web experiences to the proliferation of generative AI by web developers.
If I were a teacher and Claude my student, it probably wouldn’t scrape a passing grade most of the time.
And that makes sense, right? By the very nature of how these things work, they produce mathematically average outputs, they don’t know what good code actually looks like.
To someone ignorant of the work, it looks fine, but an expert can quickly point out how it’s not. It’s pretty much why any expert in any field will insist their job is the only one that AI cannot replace: because it’s not really good at anything.
Letting genAI freely operate in open source spaces will inevitably result in a decline in quality, and given that genAI is itself trained on those same open source projects, it’ll only become worse with time.
The collapse of community
In my work, we’ve seen a steady increase in ‘drive-by’ issues and pull requests, obviously generated by AI, created by accounts with no prior history of interacting with us and no evidence that they’ve ever used our code before.
Many of these accounts do the same thing across hundreds of high-profile code repos, clearly operated by someone seeking to inflate their resumé or GitHub contribution graph without putting in any actual effort. Their behaviour is basically no different to email spammers.
They do not care if your project is good. They don’t care about how functional, useable, or well designed your project is. They don’t care about the roadmap or the bigger picture. They don’t care about the users or the maintainers. They are not part of your project’s community and they never will be.
What they can do—if you choose to tolerate their presence—is push away the people who do care.
Even amongst people who aren’t committing ‘drive-by’ slop, merging genAI code isn’t elevating the work of an algorithm to that of a human—it’s telling the humans that their work is of as little worth to you as the algorithm’s.
Answering a bug report with an automated response or implementing a feature request with zero feedback or iteration isn’t being ‘efficient’—it’s telling your users that you don’t really care about them.
Generally, the only people who don’t get a bad taste from rampant AI usage are AI-boosters themselves, creating yet another death spiral where no one cares about anything and no one chooses to involve themselves in the process.
This is already happening
My opinion here isn’t exactly a unique one. Hell, a new post about how genAI is sloppifying an open source project, splintering its community, damaging its long-term sustainability, or just causing people to quit the industry crosses my desk almost every day (here’s today’s one).
cURL killed their bug bounty program because of an overwhelming number of genAI pull requests. Godot rewrote their contribution policies because being even slightly permissive of genAI caused such a large backlash.
Closer to my home in civic technology, the US Web Design System has been compromised by AI bros who can’t stop breaking the law and the NHS put walls up around the majority of their repos because of concerns around Anthropic’s Mythos exploiting issues.
Incentives to contribute are already disappearing; communities are already fracturing; and open source is already suffering.
We need to stop being agnostic, draw a line, and begin rejecting any genAI involvement in open source before it’s actually too late.