The short-term damage that AI will do to open-source projects seems obvious and has already started. Interest wanes; human interaction dwindles; existing projects stall; fewer new projects start. It’s probably already too late to do anything about that.
Once ouroborus manifests — the snake begins eating its own tail — then what the above author describes as “staleness” (i.e. model collapse) sets in. Also obvious and predictable. That will compromise whatever quality currently exists.
Where does that leave open-source? Is there a path forward? If not, is there even a path back to where we are now — or will that bridge have been burnt?
The path to starting a new open-source project was already pretty precarious. Warm fuzzies were the best you could really expect. Now AIs simply scrape your repo and assimilate your work without credit or compensation, and lazy programmers use AI to prompt a facsimile of your work into existence on-demand and without your knowledge. Your open-source work is stolen by American corporations and turned into an income stream, and you get nothing. When visits to your project/repo by humans drop to approximately zero… why bother?
Roll forward a couple of years, and all the charts will show a massive collapse in the number of open-source projects being hosted on public-facing — scrapable — infrastructure. Why bother starting a new open-source project when the rot has already set in?
Your chance of starting a ‘successful’ OS project at this point, and ‘making it big’, will have dropped from approximately zero (where it has been for years), to absolute zero. You can’t build a community around your code when AI is intercepting the eyeballs.
Can open-source even exist once AI has finished gutting the community?
I actually take the exact opposite viewpoint. AI has freed us from the time-triage concerns of building tools. You no longer need to decide if you want to build something, the only question is if you use that system to build in private or publicly.
For example, I built a (mostly) fully-fleshed-out discord bot in a couple weekends, which is available for public consumption, almost entirely via Qwen3.6-27b, Qwen3.8-27b and Qwen3.8-flash-next. It got a minor amount of help from Grok 4.5, for some of the higher level architectural things, but for the most part, it just workedTM
What’s better is I have the ability to deliver features, fix bugs and do what I need, from the helm of a sailboat, on the trail, at the gym, etc… I am no longer fastened to my desk while working, and I am no longer required to provide 100% focus for 100% of the development time, meaning I can do the things I really want to do in life.
But to refocus on the assertion that LLMs will replace all open source projects: Just because AI can deliver anything to you, doesn’t mean it should. And if there’s one thing that LLMs are good at, it’s reasoning. LLMs will understand that if someone else did the work already, it’s far easier to simply re-use the existing work, LLM-derived or not. And if it doesn’t do exactly what is required, or there’s a bug, the LLM can fix it, and can deliver the results back upstream for everyone to benefit.
I understand the fears, but this really is a printing-press level innovation, or put another way, this is on the level of the Apple II, not the DMCA or HDCP.
That’s not what the article was about, or what the central question of this thread is about.
It’s clearly a fact that AI tools can help develop software (both open-source and closed-source). If we ignore the issues of code quality and maintainability, I doubt there are any devs left on the planet that would deny that AI can be useful in developing software.
I also think that it is pretty obvious that the developers that use AI tools the most, and view them most favourably, are the ones that would rather be doing something else rather than coding.
Whilst some may argue that people who don’t actually want to write code should GTFO of the field entirely, I think there’s a bit of nuance in there:
An often-cited reason for abandonment of open-source projects is burnout. The project consumes too much time, leaving not enough to earn income elsewhere, put food on the table, or simply enjoy life. That’s legit, as far as I’m concerned.
Then there’s the developers who are more interested in the ‘architecture’ side of software development. Those that want to paint the broad strokes of their vision and aren’t that interested in getting into the weeds. I think that’s legit as well.
This is not an anti-AI thread.
I don’t understand the DCMA/HDCP reference, but yes, the recent jump in AI competence has resulted in a seismic shift. “Printing-press level” is a fair assessment, IMO.
(Coincidentally, I wrote my first programs on an Apple II circa 1979/1980. Silly little things that drew geometric shapes on the screen. Didn’t do anything ‘serious’ until the C64 a few years later.)
But it is. That’s the point.
I’ve shut down 2 of my 4 open-source projects in the last year because human traffic has dropped to zero. The other two have suffered roughly a 90% reduction in human traffic in the last couple of years… and will probably be shuttered in 2027. There’s no point me spending time and money keeping them online when the only things that visit are AI bots/scrapers. The humans have gone.
Yes, it is a fact that some open-source projects accept AI pull requests/patches. It is also a fact that many do not. My personal experience has been that AI-generated PRs are shit. Could they get better? Sure. Will they get better? Maybe. Does it matter? No.
I’m not remotely interested in wasting my time trying to understand what some (typically nit-picky) PR is supposed to be about when absolutely no effort has been made to adhere to the conventions of the project and respect prior work. I’m certainly not interested in large or rapid-fire patches coming from an account that has not been used to make even a single comment in any of the support forums. I’m not going to waste my time dealing with a potentially-infinite amount of AI slop. If humans aren’t talking then the community has already collapsed. It doesn’t matter how many requests AIs send upstream.
Sure, technically, “open-source” doesn’t need humans to be in the loop. All of the original definitions assumed it, though… and that’s how it historically has been. Technically, it is possible for open-source to “evolve” into something where humans or AIs create public repos, then AIs just iteratively talk amongst themselves trying to make the software better — while the human goes sailing or hiking or to the gym.
If we look at open-source through the lens of “a community of humans helping other humans develop software”, then what is left when most of the humans have been replaced by agents?
Now, to be clear, I’m not trying to go back in time. The recent improvement in AI has been seismic for the field of software development and I don’t think there is any going back. Unfortunately, I can’t see any place for an “open-source community” going forward, either. Developers will simply pay for tokens from corporations to help them code, instead of interacting with other humans in forums. I see that as sad and bad… but inevitable.
If ‘model collapse’ can somehow be avoided (at least by those models that focus on software development), then there is no point going public with your projects, either — it won’t improve model capabilities/evolution, and no human will ever notice.
We’ve lived with these sort of things for decades.
The BSD crowd intentionally enable their code to be used in the likes of proprietary things like Windows and macOS.
The GPL crowd meanwhile experienced tivoization of pre GPLv3 code, and the saasification of GPL code in general (prompting the less popular AGPL).
Some American tech companies have became obscenely rich off a FLOSS ecosystem that IMHO they’ve contributed relatively little to.
As to when humans lose interest, FLOSS is as popular as it’s every been today. Somehow we had a healthy enough ecosystem in the 90s and 00s with a fraction of the people.
We will survive. Even if 90% of the people today disappear, it will be fine, the current BSD communities and the old Linux community show that you don’t need a giant mass of people.
You made some software no one wants to use, and AI is the problem?
People have been making software no one wants decades before AI became a thing. It’s not an AI problem, it’s the simple fact that not all ideas are good, and it can be difficult to see one’s own ideas as anything other than great.
AI will change open-source, but I doubt that is a bad thing. At its core, open-source is about the freedom to use your computer without limitations, giving everyone the potential to write software is the spirit of open-source.
i’ve basically done the same. i cannot healthily coexist in this ecosystem when despite my attempts to attract co-maintainers over the years i’m still the sole maintainer, i’ve still no access to hardware to test common environments despite my attempts to ask, and people will go off and build entire businesses around what I’ve written w/ zero contribution back, financial or otherwise (looking at you, Ludus). I’m not dealing with AI BS if it’s just going to be me taking the brunt of it.
The whole premise is nonsense. Open source exists because someone is solving a problem for himself and sharing the results to whomever might find it useful.
Open source projects grow because someone else finds it better to just extend an existing project than starting from scratch.
AI does not change this paradigm, a free AI search to get to an existing project will always beat having to pay and babysit Claude to write a new one.
Same for extending a project, its faster and cheaper to have your machine minion create a plugin or add a feature you need to an existing project than rewriting an entire complex software from scratch
On an unrelated note, the thing that we know has been going on wrt OSS especially given the past two years advances in reasoning models has a face lol:
I’m not sure you understand. The methodology is more about human collaboration, and mentorship to an extent…not necessarily abandoning projects (even if for that particular author that was the primary context). It is a methodology I’ve used to motivate and reward good contributors who show prospect to grow, myself, for actively maintained projects. But it’s just not that useful anymore.
Correct… but not the scenario that is being discussed — so that isn’t relevant.
This is the scenario.
You have your own project. You want to add functionality (extend it). Normally you hunt down a package that does what you want, then you work out what the interface is and how it works, then you modify your own code to work with that interface. That all takes… time, and it makes your code… messy — because the interface style of the package you are importing into your project is never exactly the same as what you would have chosen had you written it yourself.
So no-one is talking about “an entire complex software from scratch”. We’re just talking about packages. Since open-source developers tend to add one feature at a time, the common, real-world scenario that you should be thinking of is adding a single package that does a single thing to an existing project. (That it may be done multiple times is also not relevant to the discussion.)
If you constrain your thinking to that — typical — scenario then I’m not sure the situation is as clear as you think it is. Most models now seem to be pretty good at “extend the following code to do X”-type tasks. They usually don’t butcher your existing code, either. Token cost is usually quite reasonable. That’s why it is becoming (if it hasn’t already become) the default.
My projects’ statistics, as well as the stats from a handful of other projects that I have contact with, suggest that 18-24 months ago downstream developers ‘suddenly’ started switching from hunting down packages in open-source repos and interacting with their maintainers, to getting that functionality some other way and no-longer interacting with maintainers. Prompting additional functionality into existence seems the most logical explanation.
Of course, if you have a rational explanation for why completely different maintainers, with completely different projects, hosted on different sites, all saw the same trend emerge at roughly the same time… then please do share.
PS: My projects were/are hosted on a VPS in Australia — so nothing that GitHub/GitLab/BitBucket/… may have done during that time should have affected my stats. The other projects I know of that saw the same trend were on GitHub (mainly) and GitLab. Point: The observed trend is not isolated to a single platform.
It will, but not the open contributor part. More like the way the kernel works - torvalds doesn’t accept random people’s patches…it flows through levels of trusted people. That sort of trust / vouch system being extended everywhere seems likely to me