I'm tired of AI companies calling their models open source when they aren't

Published Nov 6, 2026, 7:00 AM EST Aggy is a veteran writer and editor in the technology and gaming space. Having served as a Managing Editor for high-traffic digital publications, alongside being an editor and consultant for over a dozen sites. Aggy's published work spans a wide and respected array of tech and gaming outlets, including WePC, Screen Rant, How-To Geek, Android Police, PC Invasion, and Try Hard Guides. Beyond editorial work, Aggy's direct experience in the tech sphere extends to app development. Aggy has published two games under Tales and is always eager to learn and do more. He also likes working on computers and researching in his spare time. He knows about Windows, Linux, Audio, Video, and much more. I keep running into the same annoying habit when a company announces a new model, calls it open source, and developers cheer. Then someone reads the license and finds a pile of conditions nobody mentioned in the press release that are the opposite of open source. Open source has a pretty clear history, but AI made it really blurry. Open source used to mean something clear AI didn't write the original rules Christine Peterson came up with the term open source back in February 1998. This was right after Netscape chose to release the source code for its web browser to the public. Bruce Perens wanted the software world to have a clear legal yardstick, so he grabbed the Debian Free Software Guidelines he'd written the year before and reworked them into the Open Source Definition. The freshly created Open Source Initiative took it over and laid out ten rules that let folks redistribute the software freely, see all the source code, build their own versions, and use it however they want without anyone being turned away. I remember hearing about developers who tried to have it both ways, like Sun's Community Source License or Microsoft's Shared Source Initiative, but the OSI approach felt like the better solution. Those licenses put limits on commercial use or demanded royalties, which broke the basic freedoms the definition rested on. The Free Software Foundation felt the same way and wrote GPLv3 specifically to stop manufacturers from locking modified software out of the devices it ran on, which is another issue for open source. Regular software gets written by people who lay out clear instructions, but AI picks up patterns from huge piles of training data, and what comes out the other end is a bunch of statistical weights, not code you can read. When AI companies put out models, they usually share just those weights plus some basic code for running the model. The real training data, the pipelines that cleaned and filtered it, the configuration decisions, and what went on inside during training all stay under wraps. On top of that, a lot of releases come with custom licenses that limit commercial use or restrict particular applications. What developers actually get is basically a black box they can run but can't fully audit. It can't be rebuilt from the ground up, and you can't really see how it came together. Slapping open source on that stretches the term way past anything it used to mean. Big releases didn't pass the test Corporate marketing doesn't water down open-source AI on its own. The open source community helped it along, in my opinion. Over the years, big-name developers, maintainers, foundations, and platforms got behind the open-weight framing and spread it around without ever asking whether it met any real open source standard. The biggest issue is that it's hard to agree on what open source means, since the term mostly rests on what the community believes and on license review, not court fights. There was nothing to enforce when giant tech companies started calling their proprietary model licenses open source. Nobody pushed back as AI got bigger, so commercial players got to borrow the trust the open source movement spent decades building while quietly slipping restrictive conditions into their distribution terms. That doesn't feel like open source to me, but it's hard to undo now. Checking the actual license agreements makes the gap plain. When Meta put out Llama 2 under a custom license with commercial user limits, when xAI released Grok-1’s weights under Apache 2.0, and when Microsoft released Phi-2 under MIT, it looked like a win for open-source AI, though the licenses differed in how open they were. But the fine print said something else. Meta's Llama Community License, for example, makes any company with more than 700 million monthly active users go get a separate commercial license, which directly clashes with the Open Source Definition's rules on free redistribution and non-discrimination. Plenty of other releases threw in acceptable use policies banning deployment in areas like law enforcement or credit scoring, or barring the use of the model's outputs to train competing systems. Those limits might make sense from a safety angle, but they break the rule that anyone can use them for anything. When the OSI formally looked at models like Llama 2, Grok, Phi-2, and Mixtral, none of them made the cut. The only ones that did were fully transparent projects like EleutherAI's Pythia, AI2's OLMo, and LLM360's Amber. The platforms and institutions that hand out AI models carry some of the blame too. Hugging Face used to lump OSI-compliant releases in with proprietary and source-available ones under the same open source tags and filters. Instead of separating genuinely open models from restricted ones, the platform went for easy searching and getting more people on board. A strict definition has its own problems Open weights are still worth something Requiring full training data release for open source status isn't as simple as it sounds. Training datasets for big models pull from massive chunks of the internet and can come with copyrighted material, which has already caused real legal trouble. Models trained on medical or financial data get personal info protected under laws like GDPR, where redistributing that data publicly is just illegal. Some training data is covered by cultural rules, like the licenses protecting Māori data in New Zealand, which limit sharing out of respect for the community's control over its own data. A strict requirement to release every bit of training data would shove open-source AI into a narrow lane of public domain text, shutting it out of fields like healthcare. There's also a real practical argument for open weights releases, even when they don't hit full open source. Publishing model weights lets developers run models on their own machines, fine-tune them, make them run faster, and check them for security problems. That's a big step up from a world where everything is a closed API run by a handful of companies. Demanding textbook open source compliance as the only acceptable standard means you'll push AI projects to go fully closed. There is no easy answer for open source I don't have a clean fix to offer. A strict definition protects the meaning of the term, but it can also shut out models trained on private or culturally protected data, and that's a horrible cost. I also don't like loose labels because they let companies borrow trust they haven't earned. Most releases will sit somewhere in the middle, and that's fine as long as people say so plainly. If you're choosing a model to build on, read the license before you read the announcement. It takes ten minutes, and it can save you months of rework.