I've been thinking about how AI companies obtained the legal right to train models on code and other creative work published online. Public repositories and similar platforms were essential to building today's models, but I don't remember explicitly agreeing—when I began publishing code years ago—that it could be used to train commercial AI systems.
These models can sometimes generate substantial portions of copyrighted code, even though the original training data is stored internally as learned patterns rather than as a conventional collection of files. If the training material includes permissively licensed code, such as MIT-licensed projects, shouldn't the resulting use still preserve requirements like attribution? How do copyright law, open-source licenses, platform terms, and the fair-use argument apply here?
5 Answers
Public visibility and copyright permission aren’t the same thing. Companies often rely on the fact that material can be accessed without a password, along with broad platform terms that grant the hosting service permission to process, index, and improve its products. Those terms may support certain uses by the platform, but they don’t automatically settle whether large-scale commercial model training is lawful or whether every downstream user is covered.
The main legal distinction is between learning from a work and reproducing it. A person can read books or study public code and then create something new without copying protected passages. AI companies argue that training works similarly: the model learns statistical relationships rather than distributing the original files. Critics respond that the scale, commercial purpose, and ability to reproduce memorized passages make the comparison incomplete. Courts are still sorting out that difference.
The simple answer is that companies generally did not obtain individual permission from every creator. They argue that training on publicly accessible material is protected by fair use or a similar exception, but that position is still being challenged in court. The legality depends on factors such as how the data was obtained, whether the use is transformative, and whether the model reproduces protected material in its outputs.
So far, this has largely been a legal gray area rather than a clearly granted permission. Companies trained models first and defended the practice using fair-use arguments, public-access theories, and platform terms. Some also changed their terms to explicitly authorize broader data use. If lawmakers or courts decide that this scale of commercial training falls outside existing exceptions, the rules may change, but there is no universal opt-out or settled standard yet.
Open-source licenses usually govern how the licensed code itself may be copied, modified, and redistributed. Whether training a model counts as copying or distributing that code is one of the unresolved issues. If a model outputs a substantial, recognizable portion of a project, that creates a much stronger copyright or license-compliance concern than an output that merely reflects general programming patterns.
That helps clarify the distinction. My concern is that publishing something publicly may allow people to read and learn from it, but it doesn’t necessarily feel like consent for a company to ingest it at massive scale and sell access to a model built from it.