Training Data Rights

Before products carried AI features, customer material ended inside your own servers. Deciding where it was stored and who could see it was the whole job.

Attaching an outside model erases that boundary. The month after Pagedeck shipped meeting-note summaries, a question Sofia received from an enterprise customer's legal team made that visible: confirm in writing whether our documents are used to train that model. Neither Grace nor Theo could answer. Checking, they found the tier they were on carried a clause permitting inputs to be used for model improvement. Meeting notes from every customer with summaries enabled had been travelling that path for four months, and nothing in the customer contract said so.

Training data rights is the question of what rights attach to that path. The moment material leaves the company, copyright, contract terms and privacy rules all engage at once, and none of that is visible from the screen where the feature was built.

The first thing to separate is that there are two directions, outbound and inbound, and different rights attach to each.

Outbound is your customers' material being used to train an outside model. What breaks here is a promise to the customer: they handed documents to you, not to a third company. Three things get checked.

1. The model provider's terms per tier. The same model from the same company differs by tier, and training use is sometimes the default.
2. Retention period and region. Even where training use is off, inputs are frequently kept for days or months for abuse monitoring, and which country that server sits in creates a separate question.
3. Your own contract wording. The data processing agreement (DPA), which sets the scope and responsibility of processing on someone's behalf, needs that model company listed as a sub-processor, with training use expressly excluded.

Inbound is material the model learned from other people showing up in your product's output. What engages here is somebody else's copyright. Summarizing a meeting note carries low exposure, because the input is the customer's own document. A feature that produces prose, images or code from a blank prompt draws its output from what the model learned, which leaves room for results close to an original.

Rights do not come in one layer either. The copyright of whoever made the material, the terms of service of wherever it was published, and privacy rules that apply when information about people is mixed in all engage separately. Clearing one of the three settles nothing about the other two.

The work in practice runs in the same order every time: draw where data flows on one page, check the tier and settings at every company it passes through, align the contract wording, reflect what customers need told in the terms and privacy notice, and keep records of inputs and outputs. That last item is the only basis you will have for establishing what actually moved if a problem surfaces later.

The most frequent place is an enterprise customer's adoption review. Any product with AI features gets asked about training use, the deal stops when the answer stalls, and the answer only exists if it was settled while the feature was being built.

It also belongs in contracting with a model provider. Choosing on price and latency alone leaves the terms to bite later. A company that will write "no training use" into a contract carries different weight in front of enterprise buyers than one that only says so in published terms.

Pre-launch review of a new AI feature has a slot for it. Legislation in the shape of the AI Act, which grades uses of AI by risk level and attaches different duties to each, asks what material a system was built on as a record-keeping item, and skipping this review leaves nothing to write down retroactively.

It reaches internal policy as well. An employee pasting company material into an outside chatbot creates the same problem independently of the product, so which tools may be used at which tier, and what never gets pasted, get written down as a list.

AI bias, the other issue in this family, looks at the same material from a different angle. Bias is the problem of skewed material producing outcomes that disadvantage a group; rights is the question of whether you were entitled to use it. Material gathered with full permission can still be skewed, and a balanced dataset can still be one nobody had the right to collect.

"Confirm in writing whether our documents are used to train the model."

The question an enterprise legal team asks before adoption, every time.

"The default tier says inputs are used for training. Let's move up and get it in the contract."

The same model carries different terms by tier.

"This feature's output only comes from our customers' own documents, so copyright exposure is low."

Separating inbound exposure feature by feature.

Grace changed three things in the week that question arrived: moved to a model tier with no training use, listed the provider as a sub-processor in the data processing agreement, and stated on the summary screen whose model the feature runs on.

One thing could not be changed: four months of material that had already left. Grace told the fourteen affected customers before anyone asked. Saying nothing is quieter, but when it surfaces in a later contract or audit, the months of silence become the bigger problem. None of the fourteen cancelled, which is the usual outcome when the disclosure arrives before the discovery does.

The most common belief is that a paid plan settles it, since paying suggests different treatment. Terms are not set by the plan, though. They differ by tier inside the same company, they differ again on one account setting, and features that adapt a model on your own material usually carry separate terms altogether. What needs reading is not the price page but the documentation for the specific path in use. A company that answered a customer on the strength of an upgrade alone can discover six months later that one setting was left at its default.

The other one is that publicly available material is free to use, since anyone can see it. Public means visible, not released. Text and images on the web usually carry copyright exactly as they are, and the terms of service where they were published often prohibit automated collection separately. Either one is enough to create a problem, and information about people adds a further layer of privacy rules on top.