Open weights are not the same as open-source AI

What “open” lets you do

A proprietary AI system is intellectual property controlled by the company that develops or sells it. The company may keep the model, training data and training process private, while giving customers access through an API. ChatGPT is one example. A proprietary product can be free to use; the owner still controls its use, distribution and modification.

Open-source AI is meant to offer more: users should be able to use the system for any purpose, study how it works, modify it and share it. The Open Source Initiative’s definition sets out those rights. In practice, genuine examples are rare, and many products called open come with limits.

Mark Zuckerberg called Meta’s Llama 3.1 the first frontier-level open-source AI model. But the 2026 International AI Safety Report said its license contained “restrictive terms” that meant it could not be considered truly open.

Open weights occupy the middle ground. The weights are values and parameters learned during training and used by a model to choose what to do next. Publishing them lets users download a model and run it on their own servers. That offers more control, but does not necessarily reveal the algorithms or the training data behind the model.

The license is part of the product

Many tools advertised as open are distributed under permissive licenses, which generally allow free use subject to limited conditions, such as crediting the creators.

These are not interchangeable arrangements. A company evaluating an AI platform needs to know what it can inspect, change and deploy—and what remains under someone else’s control.

More control, more responsibility

Proprietary tools are not automatically worse, and open ones are not automatically better. Proprietary platforms can be quicker and easier to adopt, and their developers typically provide support. The strongest proprietary models still outperform open-source models on most raw-performance benchmarks.

I think the useful question is not which label sounds most reassuring, but what control the organization actually needs. As AI takes on more work and receives more data—including potentially sensitive personal or financial information—the cost of choosing the wrong license is not just technical. It can limit how responsibly a business manages its data and how independent it remains from its provider.