LLMs for Open Source maintenance: a cautious case

When ChatGPT appeared on the scene I was very annoyed at all the hype surrounding it. Since I’m working in the fast moving and low margin business of communication and campaigning agencies I’m surrounded by people eager to jump on the hype train when a tool promises to lessen the workload and take stuff from everyone’s plate.

These discussions coupled with the fact that the training of these tools required unfathomable amounts of stealing were the reason for a big reluctance on my part when trying them out. I’m using the word stealing here on purpose, since that’s exactly the crime Aaron Swartz was accused of by the attorney’s office of the district of Massachusetts. It’s frustrating that some people can get away with the same crime when it is so much bigger. For example, OpenAI and Anthropic downloaded much more data than Aaron ever did.

A somewhat related thing happened with the too-big-to-fail banks: There, the people at the top were even compensated with golden parachutes at the end. LLM companies seem to be above accountability too.

Despite all this, I have slowly started integrating these tools into my workflows. I don’t remember the exact point in time, but since some time in 2025 my opinions on their utility has started to change. At the beginning, I always removed the attribution and took great care to write and rewrite the code myself, only using the LLMs for inspiration and maybe to generate integration tests. More and more I have to admit that they are useful, especially in time constrained projects with a clear focus and purpose.

Last month I fixed and/or closed all open issues in the django-tree-queries repository with the help of Claude Code. Is that a good thing? It could be argued I should have done the work myself. But I wouldn’t have — I have other things I want to do with my time. I don’t want to (always) work on Open Source software in the evening. I definitely also have leaned heavily on LLMs when working on django-prose-editor.

Is faster better?

We can produce more code, more features and close tickets faster than before. In my experience the speed up isn’t as big as some people may want us to believe, but it’s there. And contrary to what people in my LinkedIn feed say, that’s not an obviously good thing. Is it a race to the bottom where we drown in LLM-generated slop in quantities impossible to maintain? It doesn’t feel like that — but it’s a race that could go both ways. Throwaway code can be thrown away though, and well tested code does what the tests say, which is good enough according to my rules for releasing open source software.

Speaking as someone who has put more into the training set than they’ve taken out so far, I don’t feel all that bad using the tools. Coding agents can already be run locally with reasonable hardware requirements, at least during inference, which is where the ongoing cost sits. Maybe using them is still rationalization. But contribution and profit needing to stay in some rough balance feels like the right frame. Total abstinence isn’t the only ethical choice we have.

Community tensions

What makes me less comfortable is how communities are reacting. There are real concerns within the Django world, and not just the practical one of overworked maintainers wading through hastily generated patches that don’t actually fix anything. The deeper worry is about the communal nature of contribution: that working on Django is supposed to be a learning experience, a way into the community, and that using an LLM as a vehicle rather than a tool hollows out that process. Reviewers end up interacting with what is essentially a facade, unable to tell whether anyone actually understood the problem. That’s a real concern and I don’t want to dismiss it.

But it maps onto a different situation from what I’ve been describing. Using Claude Code to close issues in projects I maintain and understand is not the same as using it to paper over gaps in comprehension on a ticket in someone else’s project. Whether LLM-assisted contributions to Django itself are appropriate is a difficult question; whether it’s appropriate to use them when maintaining your own software less so.

There’s also a harder tension around quality. Django’s conservatism has real value: rigorous review, minimal magic, a coherent philosophy. The ORM and template system don’t need to reinvent themselves, they work well, are still evolving while staying rock-solid for all my use cases. And reading the release notes always brings me joy. But it could be more exciting more often. Quality isn’t a strictly positive thing. Everything has costs. It’s not great if the price of the bar is that legitimate bugs sit open for years because nobody has a few evenings to spend on them. It happened with django-tree-queries before I went through it with Claude Code. I think the bar for contributing to Django is too high. I would value a little more motion and a little less stability, even as someone running dozens of Django websites and apps.

Then there’s the pile-on dynamic that plays out on Mastodon and GitHub. When the Harfbuzz and chardet maintainers disclosed LLM usage, the reaction from some corners was something to behold. People expressing what amounted to personal grievance over tooling choices in projects they may not even use. There’s a particular kind of entitlement in telling a maintainer – who is keeping software alive, possibly even in their spare time – that the way they choose to do that work is an affront. Open source is a gift, whether paid or not, and nobody has to accept it, but disclosing your tooling isn’t an invitation for complaints. The ethical concerns about training data, resource use and other negative externalities are legitimate and worth raising. Performative outrage directed at individual maintainers is not the same thing.

I don’t have an easy conclusion. The tools are useful, the ethics are murky, and communities are still figuring out how to respond. A cautious, honest use of them feels better to me than the alternatives.