The AI industry frequently labels models "open source" when they are, at most, "open-weight" — a distinction that matters because it determines whether users can merely run a finished model or actually inspect, reproduce, alter and redistribute the system that created it. Open weights let developers download a model, self-host it and fine-tune it without relying on a proprietary API, offering benefits around privacy, cost and vendor lock-in, but they reveal little about how the model was actually built, since the training data, selection criteria and safety methods behind it typically remain undisclosed.

The Open Source Initiative (OSI), which stewards the Open Source Definition, acknowledges that weights alone provide only "a fraction of the information required for full accountability." Stanford HAI's James Landay argues open-weight releases amount to "open distribution" rather than genuinely open models, since outsiders cannot verify data sources, copyright issues, underrepresented languages or benchmark contamination. The OSI's own Open Source AI Definition (OSAID 1.0) has drawn criticism from figures including Bruce Perens, Luca Antiga, Bradley Kuhn and Richard Fontana, who argue it lets companies claim "open source" status without disclosing training data, effectively enabling "openwashing."