The Big AI Terms Everyone Is Getting Wrong: Open Source, Open Weight And Proprietary

Open source AI is everywhere right now. The problem is that a lot of what gets called “open source” isn’t actually open source at all.

Terms like proprietary, open source and open weight are increasingly used to describe AI models and platforms, and they’re frequently confused or used interchangeably.

The distinction matters because these labels determine who controls the technology, what you can see inside it, whether you can modify it, and what you’re legally allowed to do with it.

For businesses choosing AI models, those differences can affect everything from cost and customization to data control, transparency and long-term dependence on a technology provider.

So what do proprietary, open source and open weight AI actually mean? Here’s a simple guide to the differences.

What Is Proprietary AI?

Proprietary AI refers to algorithms and tools that are owned as private IP by the corporations that develop or sell them. In many cases, the developers keep the models, the training data, and details of the mechanics of the training locked away on their own servers, and let their customers access it through APIs. ChatGPT is the most obvious example of this. It's sometimes referred to as closed-source AI, and it effectively means the IP is private property. If it’s for sale, you can use it if you comply with the owner’s terms governing how it can be used. Proprietary software may even be available in a free version (again, ChatGPT is the best example), but use, distribution and modification are still controlled by the owner.

What Is Open Source AI?

Open source is the most transparent option, but in the world of AI, true examples are rare.

The precise meaning of open source is that the source code is freely available and can be used by anyone, in any way they like.

In reality, a lot of software described as open source, particularly in AI, is made available under a variety of licenses that do control what can be done with it.

For example, Mark Zuckerberg described Meta’s Llama 3.1 as the first “frontier-level open source AI.” However, the 2026 International AI Safety Report points out that the license Llama was released under had such “restrictive license conditions” that it can’t really be called open source.

According to the Open Source Initiative, a truly open-source AI system must let users use the system for any purpose, study how it works, modify it and share it.

What Is Open Weight?

Open weight is a middle ground where many AI applications promoted as open source really belong. Weights are data that can be thought of as values, settings and parameters that the AI works out during training and uses to decide which choices to make.

Making the weights public means users can download the model and run it on their own server, bringing it more closely under their control. It doesn’t mean it’s entirely open to them, though; the source code containing the algorithms as well as the training data, which tells us what it knows, could still be locked.

Why Licenses Matter

Software described as open source is often released under one of a number of “permissive” licenses, which, broadly speaking, give you, the user, permission to use it how you like, without paying.

Typically, they contain some minor conditions, like having to credit the original authors. Frequently used licenses include Apache 2.0, which OpenAI’s open-weight GPT-OSS models are released under, and the MIT Open Source License, used by Meta’s React framework, as well as AI models including DeepSeek and Microsoft’s Phi-4.

Developers sometimes create their own custom licenses too, as was the case with Meta’s Llama 3.1 release. Its license compelled users with more than 700 million monthly users to get permission before using it, and also temporarily made it unavailable to businesses in the E.U.

If you’re buying AI tools and platforms, choice of license can have a big impact on your ability to deliver AI in a way that’s trustworthy and accountable. It will dictate how much insight you have into what you build, and what still remains a black box.

That isn’t to say proprietary equals bad and open source equals good. Proprietary tools and platforms are often quicker and simpler to get started with. Support is generally provided by the developer, who has a strong financial incentive to keep you using it. And the most powerful proprietary frontier models still outperform open-source models on most raw power benchmarks.

But as AI capability within an organization grows, and more data, potentially including sensitive personal or financial data, is exposed to AI, serious thought should be given to the impact on data stewardship and business autonomy of the decisions you make here.