The second cost is the work that trained these models. Some of that work is mine.
I remember the coffee. Costco’s Mayorga Cubano blend, nut milk, no sugar. I was in my office chair, and the first pink light of the morning was catching the top of Stringer’s Ridge across the river, and I typed a prompt asking an agent about a passage from a book I wrote, Designing Elixir Systems with OTP. It answered. It knew the book. I couldn’t tell you now which page I asked about. Some page, let’s call it 127. What I can tell you is that I broke into a cold sweat, and that something turned over in my gut, because I understood in that second that my book was inside the machine and no one had ever asked me. That was the moment the cost stopped being an argument and became a feeling.
Let me be fair about how it got there, because the mechanism is not evil on its face. A model learns to write the way a person does, by reading enormous amounts of what other people wrote. That is not theft the way a stripped car is theft. It is closer to how I learned to write Elixir, by reading thousands of other people’s programs until the patterns lived in my hands. And the reading is exactly what makes these tools good. A model that had read nothing would help no one. When an agent writes clean Elixir, it is because it studied ours.
But there is a difference between a student who reads your book and a company that ingests it, at industrial scale, to build a product it will sell back to you. My book was not a gift to a machine. It was work, edited and rewritten and paid for, and it went into a training set without a conversation. I am not the only one. This has moved through the courts for two years. The New York Times is suing OpenAI and Microsoft. Anthropic agreed to pay about 1.5 billion dollars to settle a class action from authors whose books were used without permission, roughly 3,000 dollars a title across half a million books. You do not need a final verdict to see the principle underneath. If a person’s work made your model more valuable, they are owed something.
Here is the part that actually keeps me up, though, because the books are not the worst of it for me. The code is.
Every public repository I ever pushed is on GitHub. Every example I wrote to teach a pattern, every small library, every rough experiment I left in the open so a student could learn from it, all of it was scraped to train the coding models. I wrote that code to bring people into this craft. It now lives inside the tool that does their jobs. My open source did not just train something that competes with me. It trained something that competes with my customers, the developers I am hired to build up.
The developers noticed, and they sued too. In 2022 a group of them took GitHub, Microsoft, and OpenAI to court over exactly this. Here is the part worth sitting with. The authors got a settlement. The coders mostly got dismissed. Judge after judge threw the claims out, and the case now hangs on a narrow thread about whether the tool honored the open-source licenses attached to the code. That is the quiet cruelty of the code story. The terms were never a mystery. They were written down, in the repository, in a file named LICENSE. They spelled out attribution, and share-alike, and the other conditions the whole open-source bargain runs on. The models read the code and skipped the license, and the courts, so far, have mostly shrugged.
People will tell you fair compensation is impossible here, that you cannot price one book or one repo inside a trillion tokens. For the books, maybe that is genuinely hard, though the music industry built collective licensing a century ago and still manages to pay a songwriter when a bar plays their song. For the code, it is not even hard. Fair was already written down. The licenses said what was required before any of this existed. Nobody had to invent a system. They had to read the file they were already reading, and honor it.
🎯 Join Groxio's Newsletter
Weekly lessons on Elixir, system design, and AI-assisted development — plus stories from our training and mentoring sessions.
We respect your privacy. No spam, unsubscribe anytime.
Let me make their case for them, because they are not cartoon villains. The law was genuinely unsettled. Calling machine training a kind of reading was a defensible argument, and thoughtful people made it in good faith. No clean marketplace existed to license all of this even if a company wanted to. And “it’s public on GitHub, so it’s fair to learn from” is a sentence I have said myself, about other people’s code, without blinking. The failure is not that the question was easy. It is that a handful of companies resolved a hard question in their own favor, at the scale of every book and every repository at once, and left the rest of us to argue about it after the fact.
Which brings me to my own hands, because I am not clean here. I buy fewer books than I used to. When I research now, I read the model’s synthesis instead of the six sources it stood on, and it is faster and genuinely better, and the authors of those six sources get nothing from me. I write code with agents trained on scraped repositories, probably including my own. And I teach with the very tools that swallowed what I spent twenty years making. I am the author who stopped paying authors, the coder using everyone’s taken code, and the teacher whose materials helped train the thing now hollowing out the students. I don’t have a clean way to hold all of that. I just know that it’s true.
So who pays? The midlist novelist whose sales quietly thin. The illustrator whose style you can now summon by name, on demand, for free, a living person who never agreed to become a preset. The maintainer who gave years to a library and got a line in a changelog and nothing else. And, closer to home for me, the working developer whose craft was fed into the machine that is now named as the reason to stop hiring people like them. If you build with these tools, and I do, then you are building with work that was taken from people who were never asked. You didn’t do the taking. But now you know it happened, and knowing is the part of the bill this article hands to you.
Hold it up to the test we set for ourselves in Pay Your Own Way, the one that asks whether a technology helps people flourish. This one flourished the readers and starved the writers who fed it. It lifted the people who use the code and thinned the ranks of the people who wrote it. I can see the good. I use the good every day. But I can also see who paid for it, and I can see that it did not have to be built this way. We have to do better.
Next in this series, the third cost, the one that has been circling this whole piece: what happens to the work itself, and to the people whose on-ramp the machine just pulled up behind it.
If your work is in one of these models, and if you write or draw or ship code in the open it almost certainly is, I would like to hear how you’re thinking about it. Come find me. Groxio is at grox.io.
Sources
- New York Times v. OpenAI and Microsoft, core copyright claims allowed to proceed (2025): Axios
- Anthropic’s ~$1.5 billion settlement with authors whose books were used without permission, roughly $3,000 per title across an estimated half million books (Bartz v. Anthropic): NPR
- Doe v. GitHub (filed 2022 against GitHub, Microsoft, OpenAI over Copilot): most claims dismissed, case surviving on open-source license and breach-of-contract grounds: Developer Tech · Joseph Saveri Law Firm case page
- GitHub Copilot trained on public repositories including copyleft-licensed code (GPL, LGPL, AGPL, and others), reproducing snippets without required attribution or license notices: FOSSA
🛠 Invest in the People Behind the Craft
Bruce's series asks what we owe the people whose work makes AI useful, and how we keep training the next generation of developers. Groxio runs private training and ongoing advisory for engineering teams using AI with Elixir, Phoenix, OTP, LiveView, Ecto, Ash, and Postgres. Start with a diagnostic conversation about your team's codebase, mentorship, and engineering judgment.
— Bruce