Do We Own the Rights to a Generative Model Trained with Our Own Data?
Dear Will & AiME,
We're training a generative model using our own customer and operational data. If we improve it in-house, do we own the rights to the model?
— Head of Product in Minneapolis
Short answer💡
Not automatically. You typically don’t own the underlying model, but you can assert rights over your data, fine-tuning process, and resulting improvements—if the base model’s license, your contracts, and your documentation support it.
Dear Head of Product in Minneapolis,
As more companies move from experimenting with AI to building customized tools in-house, a new kind of IP asset is emerging: the fine-tuned model. You have your base model—open-source or commercially licensed—and you're feeding it proprietary business data, adjusting weights, and optimizing outputs. The result feels like yours. But do you actually own it?
The short answer is: not automatically. But there are smart, proactive steps you can take to put yourself in the best possible legal position.
Who Owns What in a Fine-Tuned AI Model?
Let's unpack what's really going on when you fine-tune a model:
The base model isn't yours. Whether it's open-source (like LLaMA or Mistral) or licensed (like Claude or GPT), the core weights and architecture are owned by someone else. You're building on top of that.
Your data is yours, assuming it was lawfully collected, not subject to third-party restrictions, and not exposing you to regulatory or contractual risks when used for training.
The training processis also yours, but only if you control it and document how you've modified the model. This includes prompt tuning, parameter adjustments, and performance refinements.
Ownership of the output model depends on how all three of these layers interact and how well you've documented and structured them.
What You Can Do to Strengthen Your IP Position
You may not be able to own the base model, but you can absolutely stake a claim over your value-add. Here's how:
Review the Base Model License
Before anything else, read the license. Some open-source models allow commercial use with few restrictions. Others require you to share your modifications, restrict downstream licensing, or forbid use in certain sectors. If you're not clear on your rights, you're building on shaky ground.
Treat Your Training Process Like an IP Asset
The way you fine-tune the model—your data preparation, training methods, prompt engineering, and output curation—can all represent proprietary know-how. That know-how may be protectable as a trade secret, and in some cases, elements of your system or method may be eligible for patent protection.
Use Contracts to Cement Ownership
If any external vendors, contractors, or platforms are involved in the fine-tuning process, make sure you have agreements that assign all rights to you. Don't assume "work-for-hire" applies unless your contract clearly states it.
Document Everything
Maintain records of the training process—when and how models were modified, what datasets were used, who was involved, what architectural changes (if any) were made. This not only supports your ownership claims, it positions you to defend or license your work down the road.
Think About Control, Not Just Ownership
In some cases, your best asset isn't a registered right - it's your ability to control how your model is accessed, used, and distributed. Combine legal rights with technical restrictions (e.g., API gating, access logs, watermarking) to preserve the business value of your model.
Companies that invest in fine-tuning AI models are building real enterprise value, but that value can erode quickly if legal ownership and usage rights aren't clear. If you're doing the heavy lifting, it's worth building an IP strategy around it. You may not own the foundation, but you can own the structure you build on top of it.
-Will & AiME
Three Takeaways
Fine-tuned AI models sit on third-party foundations, but your modifications, methods, and data strategy can be protected.
Use licensing reviews, contracts, and documentation to assert control and ownership over your improvements.
Treat your model like any other IP asset - protect the layers you control and structure access to maximize value.
