Proprietary AI is AI technology controlled by an owner under terms that restrict access, copying, modification, or redistribution. A proprietary model may keep its weights, training code, or training data unavailable. What is disclosed and what you are permitted to do depend on its license and service terms.
What is a proprietary model?
A proprietary model is a model whose owner controls how others can use it. You may be able to send requests through an app or API without receiving the underlying model weights. Some proprietary models can also be licensed for deployment on a customer's infrastructure. Proprietary does not simply mean “runs in the cloud” or “costs money.”
Proprietary, open-weight, and open-source AI
| Term | What to check |
|---|---|
| Proprietary | The owner restricts permissions or access. Read the actual license and product agreement. |
| Open-weight | Model weights are available, but redistribution, commercial use, or other activities may still be restricted. |
| Open-source | Check the project's license and the materials it releases. Public weights alone do not establish unrestricted use or access to training data. |
These terms describe access and permissions. They do not establish which model is more accurate, cheaper to operate, or more private.
Does an API keep your information private?
An API is a delivery mechanism, not a privacy guarantee. A hosted provider receives the request content needed to process a call. Retention, access controls, training use, and available contractual protections depend on the provider and service. Proprietary ownership does not mean your input stays only on your device.
Before sending internal documents, check which service receives them, what your organization permits, and the terms for your account. See the data-sharing checklist.
A practical buying checklist
- Identify the model, the application around it, and who operates each.
- Read permitted uses and any restrictions on outputs, modification, and redistribution.
- Check hosting, retention, data-use settings, and deletion arrangements.
- Try the same permitted sample task across alternatives and record revision effort.
- Plan how to export your work if the product changes or closes.
Provider support and managed hosting can be useful. Restricted access can make inspection, customization, and switching harder. Neither is evidence of performance on your task.
Use the tool-selection routine to compare a bounded task, then distinguish deployment options with the local AI and cloud AI explanations.