The download finishes. A developer, hunched over a laptop in a quiet room, unzips the file. Inside are the weights for a new foundation model, a multi-billion-parameter miracle promised to the world under the banner of "open." But there’s another file in the directory, a LICENSE.txt that isn’t the familiar GPL or MIT. It’s a custom document, dense with clauses that read less like a grant of freedom and more like a user agreement for a web service. Use is restricted for certain applications. You can’t use the model to train a competitor. You must attribute the provider in a highly specific way.
This is not open source. It is a marketing operation.
The original promise of open source was control. It was the right to see the source, to modify it, to use it as you saw fit, and to share your changes. It was a legal framework built on the philosophy of user freedom. What we are seeing now in the world of artificial intelligence is the inversion of that promise. The word "open" has been co-opted to mean "source-available, with conditions." You get the compiled artifact—the model weights—but not the crucial ingredients that made it: the training data, the refinement code, the full architectural blueprint.
You’ve been handed a cake, but the recipe is a trade secret. You can eat it, and you can serve it, but you can’t learn how to bake it yourself. And if you serve too much of it, or to the wrong people, the baker’s lawyers will be in touch.
This semantic sleight of hand has enormous consequences. By cloaking proprietary models in the language of openness, a handful of giant corporations accomplish two goals. First, they harness the energy of the open-source community for free. Developers flock to these "open" models, building applications, finding bugs, and creating a vast ecosystem of tools that ultimately serves the parent company. It’s a form of unpaid quality assurance and market development.
Second, it chokes off the possibility of a truly open alternative. It creates the illusion of a level playing field while ensuring the house always wins. The capital required to train a state-of-the-art model from scratch is a formidable barrier to entry. By releasing free-to-download, but heavily restricted, models, the incumbents make it just difficult enough for independent competitors to gain traction. Why struggle to build your own community-owned model when a "good enough" version is available with a simple download, albeit one wrapped in legal tripwires?
We are witnessing a land grab for the next layer of the computing stack. The companies behind these models are not acting out of benevolence. They are establishing the platforms on which the next decade of software will be built. Their licenses are instruments of control, designed to lock developers into their ecosystem. The goal is not to foster a commons. The goal is to become the indispensable utility, the default choice, the single source of truth that sends you a bill or serves you an ad.
The fight is not about a niche disagreement over licensing terms. It is about whether the future of this technology will be owned by everyone or rented from a few. The community must look past the press releases and read the fine print. When a license reads like the terms and conditions for a social media app, it’s not granting you freedom. It’s telling you the rules of the company town you’ve just entered.
Generated by Reportify AI — Automate your team's status reports, standups, and weekly updates. Try free →