Proprietary

Proprietary means that something belongs to a single owner who holds the sole rights to it. The word comes from the Latin “proprius” and roughly means “own” or “one’s own.” With a proprietary computer program, the company alone decides who may use it, what it costs, and whether anyone is allowed to look inside. The blueprint of the program, that is, the source code, generally remains secret. The opposite of this is called open or open-source: there, anyone may read, copy, and pass on the blueprint. Well-known proprietary products include Windows, Photoshop, and the iPhone operating system.

Why companies don’t disclose their models

Building a modern AI model costs a great deal of money. The computing effort for training large models can quickly reach several hundred million dollars. On top of that come salaries for researchers who are among the best paid in the industry worldwide. Whoever invests that much doesn’t want to give away that edge. If the model stays proprietary, the company can charge money for every use.

For customers, this has advantages and disadvantages. One advantage is reliability: there is a point of contact, support, and fixed contracts. One disadvantage is dependency. If a company triples its prices or discontinues a product, you can’t simply take the blueprint and carry on yourself. Experts call this trap lock-in, meaning being locked into a single provider.

An important distinction is often confused. Proprietary is not the same as paid. There are free proprietary programs, for example many phone apps. And there is open software for which companies charge money, usually for maintenance and support. What matters is not the price but who holds the rights.

What lies behind the closed door

At its core, an AI model consists of billions of numerical values, the so-called weights. They are the result of training and contain everything the model has learned. With a proprietary model, the provider does not release these weights. The training data and the exact architecture of the model likewise usually remain secret.

You can still use the model, but only remotely. You send your request over the internet to the company’s servers. There, the model computes and sends the answer back. This connection point is called a programming interface, or API for short. You can think of it like a restaurant kitchen: you order and get the food, but you never enter the kitchen itself.

In between, there are hybrid forms. Some providers release the weights but prohibit commercial use beyond a certain size. Such models are called open-weight, but not truly open source. The term “open source” is therefore often used imprecisely in the AI field.

Proprietary AI in news and products

The best-known chatbots run on proprietary models. These include GPT from OpenAI, Gemini from Google, and Claude from Anthropic. No one outside these companies knows the exact weights. On the other side are models like Llama from Meta or the models from Mistral, whose weights can be downloaded and run on one’s own machines.

The term regularly comes up in business news when market power is at issue. Analysts ask whether proprietary providers can hold their high prices as open models keep getting better. Regulators are also interested, since secret models are harder to audit than open ones.

For companies, the decision is very practical. A clinic or a bank often may not send sensitive data to outside servers. In that case, only a model that runs in one’s own data center is an option. Whoever simply wants the best available quality, on the other hand, usually reaches for the proprietary offering.