Open source isn’t a slogan for me. It’s just how most of my weeks work.

I install packages I didn’t write. I poke around other people’s issues at half eleven because something obscure broke. Last month it was a date formatting edge case someone had already suffered through in 2019. Without that thread I’d still be guessing. When I’ve got something small that might save someone else a day, I often shove it in a public repo. That habit is older than Copilot, ChatGPT, or whatever we’re calling the current wave. It’s also a big part of why the web still feels buildable. Browsers, kernels, databases, package managers, design systems, the dull glue between services. A lot of that only exists because somebody chose to share.

So when the hot take arrives that open source is dead because models trained on public repositories, I get irritated. The repos are still there. People are still merging pull requests. What’s gone soft is the quiet assumption underneath: that putting something in public meant other humans could learn from it and build with it, not that the whole pile would be blended into a closed product and sold back by the month.

Open source isn’t dead. The free lunch is.

I’m not pretending the romantic version was pure. People shared for reputation, for ideology, because fixing a bug in public was less painful than answering the same email forever. Fine. The direction still mattered. The gift went toward other builders.

And that did build a lot of the internet. Not on its own. Companies, standards bodies, proprietary software, all of it. But strip out the shared stuff and modern delivery turns into NDAs and people rewriting the same date picker. I live on that shared layer every week. I try to put a bit back when I can: a post, a snippet, a boring tool, an honest note about what broke.

Everything is fine until someone stops maintaining it (opens in new tab) was true long before AI assistants showed up. Unpaid continuity under critical infrastructure was already a bad bet. AI didn’t invent maintainer burnout. It just found another way to take from people who were already stretched.

When GitHub Copilot landed, trained on public code, a lot of the anger wasn’t “a model looked at my repo”. It was the commercial bit. Work released under licences that assumed attribution, or sharing alike, or at least reuse at a human scale, had been turned into a paid assistant. Wired got the mood early. Evelyn Woods said it plainly:

The company has taken both free-to-use and copyrighted code and put it all in a blender in order to sell the slurry to commercial and proprietary interests. It feels like it’s laughing in the face of open source.

That’s from Wired’s piece (opens in new tab). It still lands harder than most of the legal commentary that followed.

Software Freedom Conservancy went further once Copilot became a product you paid for. They told projects to leave GitHub:

Launching a for-profit product that disrespects the FOSS community in the way Copilot does simply makes the weight of GitHub’s bad behavior too much to bear.

Full argument here: Give Up GitHub (opens in new tab). The Free Software Foundation had already called Copilot unacceptable. You can disagree with half of that and still understand the bruise. “Public on GitHub” was supposed to mean a kind of participation. It wasn’t supposed to mean free training data for a subscription.

Then the lawsuit mostly fell over on copyright. A California court threw out most of it (opens in new tab). A couple of thinner licence and contract points limped on. I don’t find that particularly comforting. Courts and culture are different machines. Maintainers can still feel used while a claim gets dismissed. Licences never really encoded “please don’t train a commercial model on this”, and the law isn’t rushing to invent that protection now.

I should say the awkward bit out loud. I use these tools. They help. They also learned from the same commons I’m worrying about. I’m not outside that loop. What I’m watching for is simpler and nastier: do people stop publishing because it feels like stocking someone else’s shop?

I had a tiny version of that hesitation recently. I was about to push a public repo I could have left private. A dull utility. Saves a day. Never going to be a product. First thought was the old one: someone else might need this. Second thought was newer and slightly sour: this will get scraped, trained on, and sold back to me inside an assistant. I published it anyway. I don’t assume every maintainer will keep doing that.

None of this started with AI.

Redis, Elastic, HashiCorp, MongoDB and a stack of others already moved away from the old “do almost anything, including competing with us on our own code as a hosted service” posture. Source available terms, SSPL, BUSL, Commons Clause style limits. Still readable. Often still runnable yourself. Not the same deal. There’s a whole graveyard of those licence shifts on rugpulls.dev (opens in new tab). Some people called them rug pulls. The companies called it survival. Looking at how cloud vendors productised open work while maintainers kept the thankless bits, I can see both arguments without needing a neat verdict.

The shape repeats. Shared work creates value. Someone eventually tries to fence the valuable bit. AI coding products are that story again, only more intimate. Cloud enclosure felt like infrastructure politics. Model training feels like your weird utility function and your weekend library getting vacuumed up, then rented back to you.

If the free lunch is ending, the ending looks like fees and blocks.

Cloudflare’s pay per crawl (opens in new tab) is the cleanest product version I’ve seen: let the crawler in free, charge it, or refuse it. HTTP 402 stops being a curiosity and becomes a price. Automattic took the refuse route as default on WordPress.com and Tumblr:

We currently block, by default, major AI platform crawlers, including ones from the biggest tech companies, and update our lists as new ones launch.

That’s from their note on protecting user choice (opens in new tab). Opt outs, conditional partnerships, robots.txt lists that keep getting longer, drafts of ai.txt that try to split “crawl” from “train”. It’s messy. It’s also the clearest signal yet that “public” no longer quietly means “training permitted”.

I’m not cheering that on without caveats. A web where every useful page needs a crawler toll gets poorer fast. Independent writers, small projects, niche docs: most of them don’t have leverage or a policy team. If the only defence is a gate, then yes, open knowledge gets harder to find. Harder to mirror. Harder for a junior developer to stumble into at midnight.

I’ve learned more from random READMEs and abandoned wikis than from official vendor docs. If those start disappearing behind tolls, or never get written because the author assumes a model will just ingest and resell the answer, we all get a bit stupider in public.

And the gates won’t be fair. Big publishers can negotiate. Big products can staff crawler policy. A lone maintainer with a good README has a robots rule and crossed fingers. “Protect the commons” can slide into “only the well resourced stay findable” without anyone intending that outcome.

That’s the version that actually worries me. Not a dramatic funeral for open source next Tuesday. A slow change of posture: locked, metered, suspicious. Generosity starts to look like a mug’s game.

We’re already getting woolly language around this. Marketing says open. The terms say not for your AI company, not for your cloud, not without a cheque. Readable is not open. Useful is not libre. If we don’t say that out loud, people will keep buying the label.

I don’t think open source vanishes. I think it splits.

A few projects stay as open as ever because the maintainers want that, or because the money somehow works. Others go source available. Dual licence. Stay open but block training crawlers and accept slower discovery. Move chat behind a membership. Stop publishing the interesting parts. From a distance that looks like death, because the buffet closed. Up close it can look like people finally pricing labour that was never free to produce.

I don’t want “enclose everything” as the answer. I also don’t want the polite fiction that nothing changed and licences were always enough. Be clearer when you publish, especially if model training is a hard no. Put something back into the projects you rely on. If you’re building AI on public work, a real relationship with that ecosystem beats a thank you page. And if you run a site, pick a view on crawlers and own the tradeoff.

I support open source because I need it, and because sharing between builders still makes better software than secrecy as a default. I also think maintainers are allowed to refuse a future where the winners never filed an issue and still trained on every closed one.

That refusal will cost things. Mirrors disappear. Datasets thin out. There are fewer public examples to learn from. Those costs are why I’m not interested in the loudest version of this argument. Anger travels well on the timeline. It’s a rubbish way to run a commons.

So I keep doing ordinary things. I use the open stuff. I contribute when I can. I write in public enough that the web isn’t only vendor docs and model output. I sponsor a dependency I actually rely on. I try to file a proper issue instead of only taking. I still publish the boring utility, knowing some of it will be scraped, because findable still beats resentful for me. I also back tools that let other people refuse. That doesn’t fix the industry. It stops me sitting on the sidelines while the rules get rewritten.

The free lunch was never free. Maintainers paid in evenings and half finished hobbies. Users paid in trust, and in breakage when a volunteer walked away. Companies underpaid for years and called it efficiency.

AI made that hard to ignore. A model can read more repositories in a night than I can in a career. Informal norms were carrying a lot of the ethics. Scale broke them.

So no, I don’t think open source is dead. I think we’re done pretending public means indefinitely harvestable. What comes next will be awkward: crawl fees, clearer nos, odd licences, funded projects, and some real loss of openness where fear wins. I’d take that awkwardness over a speech about how nothing’s wrong.

I still lean toward publishing. I get why someone chooses the gate. You can keep putting work in the open, knowing a model will train on it and some people will only meet it through a paywalled assistant. Or you can block the crawlers, even if fewer humans find you. There’s no clean answer. You’re picking which downside you can stand.

The code can stay open. The lunch was never going to stay free forever.