AI Training on Customer Data

What is AI training on customer data?

AI training on customer data means a vendor uses the information you put into its product to teach its AI models. That can include files you upload, messages you send, prompts you type and the answers the AI gives back.

A model "learns" by finding patterns in large amounts of examples. Once your data has shaped a model, it is very hard to pull it back out. The model may be shared with every other customer of that vendor.

In legal documents, the words to watch are "inputs" (what you send to the AI), "outputs" (what it sends back) and customer data (everything you or your users put into the service). A training clause tells you whether any of these can be used to build or improve models.

Training, improvement and processing are different things

Vendors use three ideas that sound alike but mean different things:

• Processing (also called inference): the AI reads your data to answer your request, then moves on. Your data does not change the model.
• Training: your data is fed into the model so the model itself changes. Future answers, including answers to other customers, may reflect what it learned.
• Improvement: a broad word that can mean fixing bugs, measuring quality, building new features or, unless the contract rules it out, training models.

A vendor can honestly say "we don't train on your data" while still using it to "improve" its AI features. That is why the exact wording matters more than the headline promise.

What it looks like in a vendor's terms

A typical clause reads like this:

We will not use Customer Content to train our generative AI models unless you opt in. You grant us the right to use de-identified Inputs, Outputs and Usage Data to operate, maintain and improve the Services, including our AI features.

• "Customer Content" is the defined term the promise protects. Check what it includes. If it covers only files and messages, your prompts or AI answers may sit outside it.
• "Generative AI models" limits the promise to one type of model. Other models, such as ones that rank, score or recommend, may still learn from your data.
• "Unless you opt in" means training is off by default. The reverse wording, "unless you opt out," means it is on until someone acts.
• "De-identified" means names and other direct identifiers are removed first. The content can still be used.
• "Improve the Services, including our AI features" is the broad permission that can cover training.

Some vendors put a short, firm promise in their main contract. Anthropic's Commercial Terms of Service say "Anthropic may not train models on Customer Content from Services."

Where the permission often hides

Training rights are not always in a section called "AI." Look in these places:

• The definition of usage data. Some vendors claim ownership of usage data. If the definition is wide, it can include parts of your activity that reveal content.
• Aggregated data and de-identified data clauses. These let the vendor use combined or stripped data for its own purposes, often with no limit on what those purposes are.
• Feedback clauses. Rating an AI answer with a thumbs up or down can count as "feedback," which many terms let the vendor use freely. OpenAI's Services Agreement, for example, lets OpenAI use feedback "without restriction or compensation."
• Separate AI terms or product addenda. The main contract may say nothing, while a linked AI page sets the real rules.
• The privacy policy. It may list "developing and training AI" as a purpose for using personal data.

Opt-in, opt-out and plan tiers

The same vendor may treat different plans differently. OpenAI's business data page says it does not use data from its business products and API for training by default, while its consumer products follow other rules. Our page on enterprise vs consumer AI terms explains this split.

Opt-outs also differ in how they work. Slack's privacy principles say it uses customer data to build machine learning models for features like emoji and channel suggestions, and that workspace owners opt out by emailing Slack. Our page on the AI training opt-out covers the details.

"Improve Our Services" Clause

An "improve our services" clause gives a vendor the right to use your data to make its product better. It is one of the oldest lines in SaaS contracts, written long before generative AI.

The problem is that "improve" has no fixed meaning. Lawyers at Morgan Lewis wrote in June 2026 that AI training rights "can extend far beyond traditional service-improvement provisions." If the contract defines "the Services" to include AI features and models, improving the Services can include training those models.

Some vendors close this gap on purpose. OpenAI's Services Agreement says it "will not use Customer Content to develop or improve the Services" unless the customer explicitly agrees.

To read one of these clauses, ask four questions:

• Which data does it cover: your content, or only usage and technical data?
• Does "the Services" include AI features or models?
• Is model training named and excluded?
• Does an aggregated or de-identified data carve-out reopen the door?

Fine-Tuning on Customer Data

Fine-tuning means taking a model that is already trained and training it a bit more on a smaller set of examples, so it gets better at one task. It is another way your data can end up inside a model.

There are two very different cases:

• Fine-tuning for you. The vendor builds a custom model from your examples that only you use. OpenAI's API data controls page, for example, says fine-tuning files are kept until you delete them and are not used to train OpenAI's models.
• Fine-tuning for everyone. The vendor tunes its shared model with data from many customers, including you. That is training by another name, and other customers benefit from what your data taught the model.

Some terms draw this line directly. Mixpanel's Supplemental Terms, as Venpo reported in June 2026, commit Mixpanel not to train or fine-tune its models on customer data except for that customer's benefit.

If a vendor fine-tunes anything with your data, ask who can use the resulting model and whether it is deleted when you leave.

A real change we caught

Venpo flagged this change in Articulate's terms: Articulate's AI training promise now covers only generative models. The promise not to train on customer data was narrowed to one type of model, which leaves room for other kinds of models to learn from that data.

It is not a one-off. Venpo has also reported that Amazon's privacy notice now says it uses data to develop generative AI, that Lovable now trains its AI on your prompts and code by default, and that Replit dropped its no-training promise for Pro customers.

Why it matters to you

Your confidential information can leave your control. Pricing models, product plans, source code and customer lists are often what people paste into AI tools. Once a shared model has learned from them, you cannot ask for that learning to be undone.

You may break promises to your own customers. Your contracts and your DPA likely say you use customer data only to provide your service. If a vendor trains on that data, you may be passing it on for a purpose you never told your customers about. Privacy laws also limit data use to the purposes you disclosed, a rule called purpose limitation.

The rules can change after you sign. Many vendors can update their terms by posting a new version. The FTC's staff warned in a February 2024 blog post that adopting AI training through a "surreptitious, retroactive amendment" may be unfair or deceptive. That post is guidance, not a rule, and it does not stop a vendor from giving notice and changing course.

It shows up in your sales deals. Security questionnaires from larger customers may ask whether your vendors train AI on shared data. A vague answer can slow a deal.

What to check

• Search the terms, AI addendum and privacy policy for "train," "machine learning," "model," "improve" and "develop."
• Check whether the no-training promise covers inputs, outputs and files, or only some of them.
• Check whether the promise covers all models or only "generative" or "large language" models.
• Find out whether training is opt-in or opt-out on your plan, and who in your company controls the setting.
• Read the aggregated, de-identified and usage data clauses for any right to use your data for "any purpose."
• Ask the vendor to put a clear no-training commitment in your order form or DPA, so it cannot change with a website update.

How Venpo helps

Venpo watches your vendors' terms, privacy policies, AI terms and DPAs, and emails you in plain English when a training clause is added, narrowed or removed. The Fine Print Scan answers one question for each vendor: does it train AI on your data?