How to Check if an AI Tool Trains on Your Data
The landing page will tell you your data is safe. The terms document will tell you what actually happens to it. Here is how to find the answer in about five minutes.
To check if an AI tool trains on your data, ignore the marketing site and open three documents in this order: the terms of service, the privacy policy, and the data processing addendum if one exists. Search each for the words train, improve, model, and retention. The answer is almost always in a single sentence, and it is almost never on the page that sells you the product.
Why the marketing page cannot answer this
Landing pages say things like enterprise-grade security and your data is encrypted. Both can be completely true while the same company also uses your inputs to train models, because encryption in transit and training on content are unrelated questions. Encryption describes how data moves. Training describes what the company is permitted to do with it once it arrives.
The second reason is that the answer usually depends on which plan you are on, and marketing pages describe the product rather than your specific contract. Free and personal tiers frequently default to training enabled. Business and enterprise tiers frequently default to disabled. A page describing the product cannot tell you which of those applies to you.
The four phrases to search for
Open the terms document and use your browser's find function. These four searches resolve the question for most tools.
Search for | What you are looking for | What a bad answer looks like |
|---|---|---|
train | An explicit statement that content is or is not used to train or improve models | "We may use your content to improve our services", where services is left undefined |
retention | How long inputs are stored, and whether that differs from training use | No retention period stated anywhere in the document |
aggregate | Whether an exception lets them keep using derived or anonymised data | "Aggregated and anonymised data" with no definition of the anonymisation applied |
opt out | Whether a setting exists, where it lives, and whether it applies retroactively | An opt-out that exists but is described only in a help article, not in the terms |
The word improve deserves particular attention. Improving our services is broad enough to cover training a model, and it is the single most common phrasing used where a plain statement would be less flattering. If a document says improve and never says train, that is a deliberate choice of words rather than an accident.
Defaults differ by plan, and that is where people get caught
The most common real-world failure is not a company behaving badly. It is someone reading the enterprise terms, correctly concluding that training is disabled, and then using the tool on a personal account because the work account was slow to provision.
Confirm which plan your account is actually on, in the billing settings, not from memory.
Confirm whether the training setting is per-account, per-workspace or per-project. All three exist, and a workspace-level setting does not protect a colleague's personal account.
Confirm whether the setting applies retroactively. Turning training off rarely removes data already collected, and documents that are silent on this generally mean it does not.
Confirm what happens through the API versus the consumer app. These are frequently governed by different terms, with API traffic more often excluded from training by default.
When the terms are genuinely explicit, that is worth something
Some providers now price data use openly rather than burying it. When Meta launched its Muse Code coding agent alongside Muse Spark 1.2 on 5 August 2026, the announcement came with a two-tier structure where the cheaper Contributor tier explicitly grants permission to train on prompts and completions, as reported at launch. The detail is covered in the Muse Code Contributor tier breakdown.
Whatever you think of the trade, an explicit price attached to a stated permission is a better position for a buyer than a vague clause with no number. When you are auditing tools, treat clarity as a positive signal in itself. Companies that intend to use your data in ways you would object to have an incentive to be imprecise about it.
How to check if an AI tool trains on your data in five minutes
Run this once per tool, and write the answer down so you never have to do it twice.
Open the terms of service and search the four phrases above. Copy the sentence you find, not your summary of it.
Open your account settings and find the training or data controls toggle. Record its current state and whether you changed it.
Note which plan the account is on and the date you checked, because terms change and your note needs a shelf life.
If the tool touches client or customer data, check whether a data processing addendum exists and whether you have signed it. An unsigned DPA protects nobody.
File all of the above in one place. A short table of tool, plan, training on or off, and date checked is the entire deliverable.
If a document is ambiguous after all that, treat it as a yes. Ambiguity in a contract written by the other party is not an oversight, and assuming the unfavourable reading costs you very little compared with being wrong in the other direction.
What to do about the tools that fail
A failed check does not automatically mean stop using the tool. It means match the tool to the data. A model that trains on inputs is fine for drafting a blog outline and unacceptable for a client's unreleased financials. Most people need one tool of each kind rather than one tool that satisfies every case.
Where the data belongs to someone else, the question stops being about your preferences. The broader vendor checklist in how to vet an AI vendor covers the contractual side, and the risks worth taking seriously puts data use in context with the other failure modes. If you are shipping a product of your own, writing a privacy policy for an AI app is the same problem from the other side of the table.
FAQ
Does turning off training delete data they already have?
Usually not. Most terms treat the setting as forward-looking only. If deletion matters, look for a separate data deletion request process, which is a different mechanism with its own timeline and often its own form.
Is API access safer than the web app?
Frequently yes, because API traffic is more often excluded from training by default and governed by separate business terms. Do not assume it, though. Check the API terms specifically rather than reasoning from the consumer policy.
What does anonymised actually mean in these documents?
It is rarely defined. Removing a name from a support ticket does not anonymise the account number in the next sentence, and source code is difficult to anonymise meaningfully at all. Treat an undefined anonymisation claim as weak protection.
Can I use a tool that trains on data for client work?
Only if your client agreement permits it. Many consulting and contractor agreements contain confidentiality clauses that a training grant would breach, and the tool's pricing has no bearing on that. Check the contract you signed before checking the tool.
How did this land?
About the author
Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.
