# Who Owns Machine Code?

> Machines learned to code from a commons that thousands of people built. Now someone has to decide who owns what they write.

Documentary on HN DOCS: https://hndocs.org/t/who-owns-machine-code/

About the Hacker News story "All AI-generated code should be Open Source" (1195 points, 636 comments): https://news.ycombinator.com/item?id=50507303

## Synopsis

In November 2026, a developer's short essay argued that every line of AI-generated code should be open source, and it became the top story on Hacker News. This episode looks inside the models to see what they actually keep from the free software they learned from. It then follows the courts from the monkey selfie and Thaler's Creativity Machine to Doe v. GitHub and the $1.5 billion Anthropic settlement. It ends with the developers who argued about it and the questions the law still hasn't answered.

## Key ideas

- AI coding models learned from four decades of free software and from Stack Overflow answers licensed under Creative Commons with an attribution requirement.
- Training squeezes code into patterns, not a library. Exact copying is rare but documented, and how often it happens depends on how a match is measured.
- In January 2025 the US Copyright Office said purely machine-made material gets no copyright and prompts alone don't give enough control. Human selection, arrangement and editing can still be protected.
- Thaler v. Perlmutter settled that a work with no human author can't be copyrighted. How much human guidance is enough is still undecided.
- The monkey selfie case only decided that animals can't sue. The human-authorship rule comes from Copyright Office guidance and from Thaler.
- Countries disagree: the UK protects computer-generated works but may repeal that, and a Beijing court protected an image made from prompts.
- A license needs copyright underneath it. Public domain is not open source, and uncopyrighted code can still be kept secret.
- Developers can document their own contribution, scan AI output for copied code, and choose an explicit license when they want their code to be open.

## Transcript (English)

September, nineteen eighty-three. At the MIT Artificial Intelligence Lab, Richard Stallman posts a message to Usenet announcing GNU, a free operating system anyone may share. Four decades of shared code followed. Now machines trained on that code are writing software of their own.
On November sixteenth, two thousand twenty-six, a short essay by the developer Jakub Jankiewicz, known as jcubic, topped Hacker News. Its claim was blunt. Every line of code written by artificial intelligence should be open source.
His first reason is a debt. The models learned from decades of free software, and from Stack Overflow answers shared under a Creative Commons license that requires attribution. Turning that into proprietary code, he writes, feels like license laundering.
His second reason is legal: machine output has no human author. He points to a famous photograph. Naruto, a crested macaque, grinning into David Slater's camera in Indonesia, two thousand eleven.
Wikimedia Commons already tags purely machine-made images as public domain. So jcubic proposes a voluntary norm. Code made mostly from prompts, barely reviewed by a human, should be released under an open license, or given away.
To weigh that, look inside the machine. Training is not a library. Billions of code files are boiled down into a fixed set of numbers, the weights. What survives is mostly pattern: idioms, interfaces, structure.
Exact strings survive only occasionally, usually text repeated thousands of times, like license headers. But in two thousand twenty-two, Professor Tim Davis of Texas A and M showed Copilot reproducing large chunks of his own sparse-matrix code.
How often? GitHub's early figure was about one suggestion in a thousand. Academic studies of other models ranged from that to one in ten. The answer depends on how long a match must be, and how it is counted.
Generation is a weighted roll of the dice. At each step the model samples the next fragment of text. The same prompt can yield different code, yet popular problems collapse into the same few answers.
That randomness matters in law. In January two thousand twenty-five, the United States Copyright Office concluded that purely machine-made material is not protected, and that prompts alone do not give enough control. Human selection, arrangement and editing still can be.
Stephen Thaler tested the extreme. He sought to register an image he said his Creativity Machine made alone. Courts refused, and in March two thousand twenty-six the Supreme Court declined to hear it. No human author, no copyright.
The monkey case, though, decided something narrower than the essay suggests. The Ninth Circuit ruled only that animals cannot sue. The human-authorship rule comes from Copyright Office guidance, and from Thaler.
Between the extremes lies a grey zone. The Office registered A Single Piece of American Cheese, built in about thirty-five rounds of AI inpainting, protecting only the human arrangement. Jason Allen's Midjourney image was refused. Critics call the line arbitrary.
Abroad, the map fractures. Britain still protects computer-generated works, though its government now favours repeal. In Beijing, a court protected a Stable Diffusion image, counting the user's prompts as intellectual investment.
Then the output. In September, the Ninth Circuit ruled in Doe versus GitHub that Copilot's suggestions are new works, not copies stripped of their copyright notices. How far that ruling reaches is still debated.
And the input. Judge William Alsup found that training on lawfully bought books, cut from their spines and scanned, was fair use. Pirated libraries were not. Anthropic settled for at least one and a half billion dollars.
On Hacker News, nearly twelve hundred points and more than six hundred comments. Many saw a contradiction. A license is a building that stands on copyright. With no copyright underneath, there is nothing for the license to hold.
Others spoke from experience. The commenter blagie, author of copyleft code, described a company cloning his project with a language model. Like rewriting Harry Potter, he wrote, with the same plot but bad dialogue. His account cannot be verified, but early memorization is documented.
Practitioners pushed back. Steering an AI means months of testing and product decisions, and humans carry the risk. One claim, that nearly all code is now machine-written, is false. Surveys this year put AI-assisted code at roughly half.
Untangled, the essay makes three claims. Machine code should be shared. It cannot be owned. Everyone should license it openly. Only the second has firm legal support, and only for purely machine-made output.
And public domain is not open source. Uncopyrighted code can still be locked away as a trade secret. If it escapes, anyone may fold it into closed products. The opposite of what copyleft authors wanted.
Some say the debt is owed elsewhere: open the models, not every output, or pay for what was taken. Meanwhile, developers can document their choices, scan for copied code, and pick an explicit license when they mean it.
How much human guidance makes an author is still undecided, and each country draws the line differently. Forty-three years after a message on Usenet, the commons Stallman started faces a new question. Who owns what it taught?

## Transcripción (es)

Septiembre de mil novecientos ochenta y tres. En el laboratorio de IA del MIT, Richard Stallman anuncia en Usenet GNU, un sistema operativo libre. Siguieron cuatro décadas de código compartido. Hoy, máquinas entrenadas con él escriben su propio software.
El dieciséis de noviembre de dos mil veintiséis, un ensayo del programador Jakub Jankiewicz, alias jcubic, encabezó Hacker News. Su tesis era tajante: toda línea de código escrita por IA debería ser abierta.
Su primer motivo es una deuda. Los modelos aprendieron de décadas de software libre y de respuestas de Stack Overflow con licencia Creative Commons que exige citar la autoría. Convertirlo en código privativo, escribe, parece blanquear licencias.
El segundo, legal: lo generado por máquina no tiene autor humano. Cita una foto famosa: el macaco Naruto ante la cámara de Slater, dos mil once.
Wikimedia Commons ya marca como dominio público las imágenes puramente artificiales. Así, jcubic propone una norma voluntaria: el código hecho casi solo con instrucciones, apenas revisado por humanos, debería llevar licencia abierta o regalarse.
Miremos dentro de la máquina. Entrenar no es archivar. Miles de millones de archivos de código se reducen a números fijos: los pesos. Queda sobre todo el patrón: modismos, interfaces, estructura.
El texto literal rara vez sobrevive, salvo lo muy repetido, como los avisos de licencia. Pero en dos mil veintidós, Tim Davis, de Texas A&M, demostró que Copilot reproducía largos fragmentos de su código.
¿Cada cuánto? GitHub dio al principio una de cada mil sugerencias. Estudios académicos, hasta una de cada diez. Depende de cuánto mida una coincidencia y de cómo se cuente.
Generar es tirar dados cargados: el modelo sortea cada fragmento. Una misma petición da código distinto, pero los problemas comunes acaban en pocas respuestas.
Ese azar importa. En enero de dos mil veinticinco, la Oficina de Copyright estadounidense concluyó que lo puramente artificial no se protege, y que no bastan las instrucciones. La selección, el orden y la edición humanos, sí.
Stephen Thaler fue al extremo: registrar una imagen hecha solo por su Creativity Machine. Ningún juez cedió; en marzo de dos mil veintiséis, el Supremo lo inadmitió. Sin autor humano, no hay copyright.
Pero el caso del mono dice menos de lo que cree el ensayo: solo que un animal no puede demandar. La autoría humana la exigen la Oficina de Copyright y Thaler.
Hay una zona gris. La Oficina registró A Single Piece of American Cheese, hecha con unos treinta y cinco retoques de IA, pero solo por su composición. La imagen de Jason Allen con Midjourney, no. Para los críticos, arbitrario.
Fuera, el mapa cambia. Reino Unido protege lo generado por ordenador, aunque quiere derogarlo. En Pekín, un tribunal amparó una imagen de Stable Diffusion: las instrucciones eran inversión intelectual.
El resultado: en Doe contra GitHub, el Noveno Circuito dictaminó en septiembre: Copilot crea obra nueva, no copias sin aviso. Su alcance se discute.
Lo que entra. El juez William Alsup falló que entrenar con libros comprados, desencuadernados y escaneados era uso legítimo. Con bibliotecas piratas, no. Anthropic pactó pagar al menos mil quinientos millones de dólares.
En Hacker News, casi mil doscientos puntos y más de seiscientos comentarios. Muchos vieron una contradicción: una licencia es un edificio sobre el copyright. Sin cimientos, se cae.
Otros hablaron por experiencia. blagie, autor de código copyleft, contó que una empresa clonó su proyecto con IA: «Como reescribir Harry Potter: misma trama, malos diálogos». No es verificable, pero los primeros modelos sí memorizaban.
Los profesionales replicaron. Guiar una IA exige meses de pruebas y decisiones de producto, y el riesgo es humano. Y la IA no escribe casi todo el código: las encuestas de este año hablan de la mitad.
Son tres ideas: el código de IA debe compartirse, no tiene dueño y todos deberían abrirlo. Solo la segunda tiene base legal, y solo en lo puramente artificial.
Dominio público no es código abierto: sin copyright, puede ser secreto comercial; si se filtra, cualquiera lo usa en productos cerrados. Lo contrario del copyleft.
Para algunos, la deuda es otra: abrir los modelos, no cada resultado, o pagar lo tomado. Mientras, los programadores pueden documentar, rastrear código copiado y fijar una licencia explícita.
Sigue abierto cuánta intervención humana hace a un autor, y cada país decide distinto. Cuarenta y tres años después de aquel mensaje, el procomún de Stallman afronta una pregunta: ¿de quién es lo que enseñó?

## Sources

- [All AI-generated code should be Open Source](https://jakub.jankiewicz.org/blog/ai-code-open-source/)
- [Hacker News discussion (636 comments)](https://news.ycombinator.com/item?id=50507303)

Footage credits are listed in the film's page. Narration and scripts are AI-generated; if you find a mistake write to the address in the About page.
