Few Thoughts on Machine Learning Agreements or AI Agreements

Short answer: AI and machine learning agreements are not standard software licenses. Three issues drive everything: how the system works and who does what, who has what rights to the trained model, and who owns the underlying software and algorithm.

Machine learning agreements and AI agreements are relatively new, and actually very interesting. Not much has been written on these agreements, so I thought I would share a few thoughts on the big issues (from the perspective of the AI or machine learning software vendor).

So here goes.

How the AI System Works and Who Does What Matters.

• Everything starts and ends with the answer to this question.
• It drives the warranties you provide, the IP ownership model, and in general the rights and obligations of the parties.
• Example: whose algorithm is being used? Is this a perpetual or term-based deal?

Who Owns the Machine Learning Model?

• Usually customers own their input and output data from software programs, but it is not that simple with AI and machine learning agreements.
• Machine learning models are generally viewed as the output from training a learning algorithm. The model is then applied to the dataset. The model may contain parts of the underlying algorithm, so ownership of the model matters less than who has what rights to it, and for how long.
• One way to solve it is to allow both parties to use the machine learning model during the term, then have the vendor destroy the model at the end. The vendor stays free to independently recreate future models, and the customer is protected against the vendor re-using its model for other customers.

What About the Underlying Software and Algorithm?

• This one is much easier.
• The vendor brings the software to the party and should own any enhancements, modifications, or changes to it (and to its algorithms).
• All AI and machine learning systems sit on top of or are embedded in a software program, so industry-standard licensing terms should apply to the underlying software.

Training Data Rights Are the Quiet Fourth Issue.

Since I first wrote this, training data has become the issue clients fight over most. The model is only as valuable as the data it learned from, so the agreement has to answer: can the vendor use the customer’s data to train or improve models, and if so, are those models used only for that customer or across the whole customer base? Vendors generally want a right to use de-identified or aggregated data to improve the service. Customers increasingly say “train on my data only for me.” Spell out the scope (this customer vs. all customers), the form (raw, de-identified, or aggregated), and whether anything survives termination. Vague “we may use data to improve our services” language is where these deals go sideways.

Who Is Liable for What the Model Outputs?

The other modern battleground is responsibility for the output. AI systems can produce wrong, biased, or infringing results, so do not let your contract promise more than the technology can deliver. Practical moves for a vendor: disclaim warranties on the accuracy or fitness of AI output, make clear the customer is responsible for human review where the stakes are high, and be careful with IP indemnities, because promising that AI-generated output infringes nobody’s rights is a big and uncertain promise right now. Tie your indemnity to your own software and algorithm, not to whatever the model happens to generate from a customer’s prompts and data.

The throughline: data and model rights are the new battleground, while the underlying software follows familiar licensing rules. When drafting your AI or machine learning agreement, remember it will not be as simple as a standard software license. Dig in and figure out the issues, so you and your customers both know what they can and cannot do with your software and the output.

For the foundational decision between a license structure and a SaaS subscription structure that underlies any AI deal, see SaaS Agreement vs. Software EULA: Which Template Do You Need?

Resources:

• Here is an article from the other side, which I think is too simplistic about who should own what. Link

Disclaimer:

This post is for informational and educational purposes only, and is not legal advice. You should hire an attorney if you need legal advice, which should be provided only after review of all relevant facts and applicable law.