“Your Data Doesn't Train Our Model.” Technically.

It is an LLM contract clause that enterprise buyers seem to take real relief in. But the risk is not in what it says. It is in what it leaves unsaid.

What you'll learn

What a “no training” clause restricts — and what it stays silent on

How made-up data can carry your know-how without using your files

Two older contract terms that quietly permit what you just banned

What to ask for instead

The clause everyone asks for

Executives often say, “The contract says they won't train on my data.” Getting that in writing is the right thing to do. But as Satya Nadella pointed out, protecting your knowledge takes more than protecting your data.

The clause usually covers your files as an input. It often says nothing about the patterns drawn from them — and patterns are the part that matters. How a company makes money is often the most valuable part of the business: the pattern is the key point.

Lawyers have a word for this. A contract is silent on anything it does not name. Silence is not protection. It is just a subject nobody wrote down. Nadella names it plainly, calling it ironic that providers impose strict terms on their own models while reserving “the right to learn from customer usage and interaction data.”

The gap: made-up data that acts real

A model can study how your work is shaped, then generate artificial examples with the same shape. None of it is your actual data, but it behaves like your data. This is synthetic data, and it is now standard practice.

85–95%

Of the usefulness of real data, achievable with well-built synthetic data for AI training

What that means

A vendor can honour “no training on your data” to the letter and still end up with a model that has learned how your business runs.

This is not an accusation. It is a gap between what the words say and what buyers hear.

NetApp | Iterate.aiYour Data Doesn't Train Our Model. Technically · 01 / 02

Two more silences worth checking

When reviewing a contract, make sure to look for two additional terms, or the lack of these terms. Get clarification on each:

“Improve the service”

Older contracts often let a vendor use your activity to improve or enhance the service. That wording was written before AI training existed — and it is broad enough to cover it.

Derivative data

Unless the contract defines this term, anything built from your data may not count as your data at all. Statistics, embeddings and summaries can all sit outside the definition.

What to ask for instead

Five questions turn a vague promise into something you can hold a vendor to.

Define derivative data — and say plainly that anything derived from our data is treated as our data.

Name synthetic data — no generating artificial data modelled on ours, for training or any other purpose.

Scope “improve the service” — so it excludes model training, tuning and evaluation.

Cover the exhaust — prompts, agent traces and corrections, with retention and deletion terms stated.

Ask where it runs — which machines, and who else queries the same model.

The version that needs no clause

Every question above exists because the model sits somewhere you do not control. Run it on your own infrastructure and most of them stop being contract problems. AIPod Mini keeps the model, the data and the learning inside the building.

The only promise that holds without a lawyer

A promise not to train on your data is worth having, but that promise might not cover replicating your knowledge. The only version that holds without a lawyer is the one where the model never leaves your building.