Who owns model weights when contractors train on a company's records?
Your company owns the model weights only if your contractor paper says so. Owning the records the model learned from doesn't carry over to the weights, and paying for the training doesn't either. You own them when the agreement names the weights as a deliverable assigned to the company, and your data access terms bar the contractor from reusing your records anywhere else. That's the core of how I approach making sure a technology company owns what it creates.
Why a standard contractor agreement can miss the weights
Most contractor agreements were written for software. They assign "code," "documentation," or "work product" to the company. Say you hire a contractor to fine-tune a model on five years of your support tickets. The weights that come out are neither code nor documentation: they're millions of numbers the training run produced.
That matters because of the default rule. Under US copyright law, what an independent contractor creates belongs to the contractor unless a signed writing assigns it. And whether copyright protects model weights at all is still unsettled, so an assignment that leans only on copyright may not reach them.
Here's what that costs you. If the contractor keeps a copy, you may have no clean way to stop them reusing it for the next client in your market. And when an investor or buyer asks you to show the company owns its core technology, "we paid for it" won't close that diligence item.
Your records, the training work, and the weights each need their own line
The usual advice stops at "the contract decides, so ask your IP lawyer." That's true, but it doesn't tell you what the contract should say. When a contractor trains on your records, three different things are in play, and your paper should handle each one on purpose.
- 01
Your records
They stay yours. Your confidentiality and data access terms decide what the contractor may do with them while they have access, and after.
- 02
The training work
Scripts, configuration files, and evaluation sets. Assign these to the company the same way you'd assign any code.
- 03
The weights
Checkpoints, adapters, and the final model. Name them as deliverables, assign them, and treat them as your confidential information, so you're covered even where copyright isn't clear.
One limit to know: if the contractor fine-tuned an open model, the weights also carry that base model's license. Your agreement can't override it, so read that license before anyone picks the model.
What if your contractor wants to reuse what they built? That's fair to negotiate. You can let them keep their general tools and know-how while the weights trained on your records stay with you. If your agreement was signed before anyone said "model," tell me about your deal.
Check your paper before the next training run
You can do most of this in one sitting with the contractor agreement and the data access terms open side by side. Each check is a line you either have or can add.
- Search for the words. Look for "model," "weights," and "checkpoint." If none appear, the assignment probably doesn't name them.
- Define the deliverables. List weights, checkpoints, adapters, training scripts, and the configuration needed to retrain, all assigned to the company.
- Limit the records. Your data is used only for your project, never to train another model, and is returned or deleted at the end with written confirmation.
- Tie delivery to the last payment. Final weights handed over, plus a written statement that no copies remain, before the final invoice is paid.
- Check privacy if the records hold personal data. Training is a new use of that data. My guide to analyzing AI under the GDPR and CCPA walks through how to test it.
Working on something like this? Tell Charlotte about it.
Tell Charlotte about your dealRead the full argument
The weights are one case of a wider rule: you own what your contractor, collaboration, and licensing paper says you own.
Charlotte Lynn Luu
Attorney & Corporate Counsel, Charlotte Corporate Counsel
Charlotte is a Georgia-licensed attorney focused on contract negotiation and risk management for startups and growth-stage technology companies. IP ownership is one of her four practice areas, covering ownership agreements, licensing, collaboration agreements, and confidentiality.
