Skip to content
Back to the blog
  • Free Software
  • Open Source
  • Inteligência Artificial
  • Forças Produtivas
  • Condições de Trabalho
  • Tecnologia e Sociedade

The Developer's Manifesto

AI, Free Software and Labor

The AI that autocompletes your code learned from the code of millions of people. Part of it was under the GPL, a license written precisely so that nothing derived from that work could become closed property. Today that learning comes back to us as a monthly subscription. I wrote the first version of this manifesto in 2025, and it treated free software almost as an idea. This version is about labor: who writes the code, who keeps the value it creates, and what a developer can stand for when AI and the cloud fence off what used to be held in common. At the end, I lay out the commitments I make and where I still fall short of them.

Why a Developer's Manifesto in 2026

Software is collective work. Almost no production system is written from scratch: it rests on libraries, compilers, kernels and tools written by thousands of people, many of whom were never paid. Together, they form the digital commons: a shared heritage nobody bought and everybody uses.

I have written about the history of this commons before. In Free Software vs Open Source I show how the movement split between freedom and efficiency, and in Open Source I show how big companies learned to use that work while giving almost nothing back. I won't repeat those posts here. What has changed since then is the reason for this manifesto, and it comes down to two things happening at once.

The first is that AI has become the newest fence around the commons. Models trained on public code sell that work back as a paid product, and appropriation has become faster and less visible. A company used to have to take a project and close off a fork. Now it only has to train a model.

The second is that developers stopped seeing themselves as workers. Many of us think of ourselves as "entrepreneurs of the self": contractors, freelancers, people negotiating alone with the market. Ricardo Antunes and Ruy Braga were already studying this process in Infoproletários back in 2009: the precarization of work in IT, including those hired as contractors (PJ, a Brazilian legal entity) instead of employees. If you see yourself as a one-person company, you miss the fact that the code you write, at your job and in open source, is labor appropriated by others.

Each feeds the other. AI speeds up appropriation, and developers lack the class consciousness to push back. Marx called the theft of common land and its conversion into modern private property primitive accumulation. What is happening to code is not the same thing, but it rhymes.

AI Trained on Open Source Code: the Newest Fence

GitHub Copilot was trained on millions of public GitHub projects. In 2022, a group of programmers sued GitHub, Microsoft and OpenAI in Doe v. GitHub. In September 2026, the U.S. Court of Appeals for the Ninth Circuit upheld the dismissal of the main claim, based on the Digital Millennium Copyright Act (DMCA), on the grounds that the model creates new works and does not "remove" authorship attribution.

This ruling deserves a careful reading. It does not say that training AI on open source code is legal. It only says that this particular DMCA provision does not apply. The breach of contract claims, that is, the license violation claims, are still pending in the trial court. The question that matters to us remains open.

Microsoft argues that training on licensed code is fair use, a use permitted without authorization. The Software Freedom Conservancy, an organization that defends copyleft, replied to the U.S. Copyright Office in 2023:

"To concede Microsoft's 'fair use' claims would be the first step in eviscerating the copyleft licenses that protect the primary commons of software source code."

In other words: accepting Microsoft's argument would hollow out the very licenses that protect the largest commons of source code we have. The SFC also points out that Microsoft acknowledges having used large amounts of copyleft code in training.

"Open Source AI" without the data

In October 2024, the Open Source Initiative published its definition of open source AI. It does not require training data to be open: it only asks for sufficient information about it. The OSI FAQ argues that "training data is not equivalent to source code". The criticism came from within. Bruce Perens, author of the original Open Source Definition, replied that training data is the source code of a model.

It is the same Free versus Open split, now applied to AI. On one side, a definition that accommodates the interests of those who train models on data they cannot or will not release. On the other, the Free Software Foundation, which argues in a draft that a machine learning system is only free if its training data and scripts are free as well.

The technology is ours

I don't think the answer is to reject AI. I use AI every day, and it was built on top of our work. That is exactly why it is ours too. The fight is not about using it or not, it is about who controls what was learned from the commons.

Given what has been happening, I argue that AI needs copyleft. If a model was trained on free code, it should be required to be open: weights, data and process. Today this is a proposal, not a license in force. The SFC set up a committee on the subject in 2022 because of Copilot, and in 2024 it concluded that no available system can be reproduced by the public. This is a political demand, and political demands are built by people standing up for them.

Relicensing and the Cloud: When the Commons Becomes a Product

AI is not the only fence. Over the last three years, widely used open source projects changed their licenses to stop others from profiting from them.

  • Terraform (HashiCorp): in August 2023 it swapped the MPL 2.0 for the Business Source License, which is not open source. In September, the Linux Foundation launched OpenTofu, a fork backed by more than 140 organizations.
  • Redis: in March 2024 it dropped the BSD license for two source-available licenses, admitting it was no longer open source. A week later Valkey was born, a fork maintained under the Linux Foundation.
  • Elastic: closed the Elasticsearch license in 2021 and, in August 2024, brought back the AGPL as an option.

I see two readings that are true at the same time.

The first is that this was appropriation from the community. These projects grew on contributions from many people outside the company. Closing the license afterwards means taking what was built in common and turning it into a private asset.

The second is that it was a symptom of cloud concentration. Redis itself justified the change by saying most of its sales went through the big cloud providers, which turned the project's investment into a commodity. In 2025, when adding the AGPLv3, CEO Rowan Trollope said cloud providers "reap the profits and control the infrastructure without proportional contributions back". That is capital itself acknowledging the extraction.

And there is a contradiction here worth facing: Valkey, the fork that "resisted" Redis closing its license, is backed by AWS, Google Cloud and Oracle. The commons resists, but the cloud is still the one paying for the resistance. Elastic went back to the AGPL after being named AWS "partner of the year". Forks prove that free code survives a closure, but they don't change who concentrates the infrastructure.

That is why the AGPL matters so much. It is the GPL adapted for software that runs as a service: whoever offers the program over a network must make the modified code available. It is copyleft's answer to the cloud, and it is no accident that Redis and Elastic went back to precisely this license. I explained the difference between copyleft and permissive licenses in GPL vs MIT.

The Invisible Work of Open Source Maintainers

On March 29, 2024, engineer Andres Freund announced that versions 5.6.0 and 5.6.1 of xz utils, a compression library shipped in nearly every Linux distribution, contained a backdoor. The flaw received the maximum severity score (10.0) and came close to reaching SSH on servers around the world. The malicious code was inserted by "Jia Tan", a contributor who spent years earning the project's trust.

The most revealing part happened two years earlier. In June 2022, under pressure from accounts demanding a new maintainer (later suspected of being part of the attack), Lasse Collin, the original maintainer, wrote on the project's mailing list that his ability to care for xz was limited by long-term mental health issues, that he had been working with Jia Tan, and closed with this:

"It's also good to keep in mind that this is an unpaid hobby project."

A project running on millions of servers was maintained as an unpaid hobby by one exhausted person. This is not an exception. In December 2021, Log4Shell exposed another 10.0 flaw in Log4j, a Java library used by companies all over the world. The official U.S. government report on the incident opens by noting that "a group of volunteers maintain Log4j" and discusses the risks of a "thinly-resourced, volunteer-based" community.

The numbers confirm the pattern. In the 2024 report by Tidelift, a company whose business is precisely paying maintainers, 60% of maintainers describe themselves as unpaid hobbyists, only 12% earn most of their living from this work, and 60% have quit or considered quitting a project. One anonymous maintainer summed it up in the report: open source has generated trillions in value, harvested by large companies that give very little back, "not even recognition, respect, or gratitude".

In Marxist terms, this is the appropriation of other people's labor: value produced by those who work and captured by those who control the infrastructure. In a job, that appropriation happens through the wage. In open source, it happens with no wage at all.

Who should pay

I don't believe in a single solution. I believe in three fronts working together:

  • Companies that profit from software pay, as an obligation and not as charity or a marketing campaign.
  • Public money funds digital infrastructure, the same way it funds roads. Germany already does this: the Sovereign Tech Agency has invested €41.1 million in 118 technologies since 2022, including OpenSSH, PHP and systemd.
  • Maintainers organize into foundations, cooperatives and unions to bargain collectively, instead of asking for donations one by one.

What I Commit to Defending

A manifesto that only diagnoses is just an analysis. This one has commitments, and they are concrete enough for me to be held accountable for them.

The first is to use copyleft by default. When I publish a project, the license is a political choice, not a repository detail. Rubrica, Gatekey and Code Forge are under GPLv3. Voxr, which runs as a service, is under AGPLv3, because that is exactly the case the AGPL exists to protect. Permissive licenses have their place, but by default, whoever uses my work to build something should give back what they built.

The second is to give back to what I use. Every dependency in my package.json has someone behind it. Giving back is not always code: it can be a well-written bug report, a documentation fix, an answer on an issue, or money, when the company I work for can pay. xz showed what happens when nobody gives back.

The third is to prefer open standards, tools and models. The open web over closed platforms, open formats over formats that lock you in, open-weight AI models when they solve the problem. Each of these choices reduces dependence on whoever controls the infrastructure.

The fourth, and the hardest, is collective organization. Recognizing myself as a worker, not as a one-person company. This takes concrete forms: unions in our field, such as SINDPD, the IT workers' union in São Paulo; free technology cooperatives, such as EITA, which builds technology alongside social movements and has already become the subject of academic research; and examples abroad, such as the Alphabet Workers Union, founded in 2021 by just over 200 Google workers and which today claims more than a thousand members. I wrote about cooperative alternatives in Platform Cooperatives and the Fediverse.

The Limits of a Manifesto

I need to be upfront about a contradiction: I use closed tools. I write much of my code with Claude Code, a proprietary product. I work on platforms I don't control. An attentive reader will notice, and it is fair that they do.

Nobody lives outside the system they criticize. A manifesto is not a purity test, it is a direction. It concerns what I control: the license of what I publish, what I give back, the choices I can make and the organizing I take part in. Waiting for a life without contradictions before acting is the most efficient way to never act at all.

I also need to be upfront about free software: on its own, it emancipates no one. Free code without organization becomes free raw material. That is what happened with Linux in the cloud, with Log4j in corporations and with GitHub code in Copilot's training. The license protects the code; what protects the people who write the code is their organization. That is why the fourth commitment weighs more than the other three.

Conclusion: Pick One Commitment

AI learned from our work, companies close their licenses when the cloud squeezes them, and the world's infrastructure depends on people maintaining software as a hobby. None of this is an accident: these are the incentives of a system that turns the commons into property whenever it can. As in Engineers of Freedom or Architects of the Prison?, the question is which side our work is on.

You don't have to adopt all four commitments at once. Pick one and apply it to your next project:

  • License: publish your next repository under the GPL or AGPL and explain why in the README.
  • Give back: pick a dependency you use every day and contribute to it this month, with code, documentation or money.
  • Openness: replace one closed tool or model with an open alternative where it does the job.
  • Organization: find out which union represents your field and talk to your colleagues about it.

The code we write is already collective. What is missing is for the decisions about it to be collective too.

References

Matheus Mierzwa

AI Enablement Lead & Frontend Tech Lead

I write about what I build and lead in practice.