“Is my data used for training?” is the question people ask, and its flaw is being one question about three different things. Text sent to a service can end up in three places that are nothing alike, and the answers are not the same.
One: training
The most discussed case and the rarest in business offerings. Free-tier terms and work-plan terms are often opposite on this point, and the difference sits in a line of the contract, not in the interface.
If the text enters training, the data cannot be retrieved or erased in any strict sense: you can delete the copy, not the effect it had on the parameters. That is why the right to erasure, applied to models, is an open problem rather than a procedure.
Two: request retention
Almost every supplier keeps requests for a stated period even when they do not train: they use them for abuse detection and support. The number is in the documentation, usually in days, and it is worth reading: thirty days on a European server and indefinite retention elsewhere are different situations for anyone who has to fill in a record of processing.
During that period the text exists in a system that is not yours, with employees who in defined cases can read it, and with the possibility that an authority in that country asks for it.
The number of days is a promise, not a law of physics
It is worth looking at what can happen to that number, because it did happen, and not to some marginal supplier. In May 2025, in the litigation brought by the New York Times, a US federal judge ordered OpenAI to preserve and segregate output log data that would otherwise have been deleted (including, as it was framed, conversations users had deleted). The obligation stood until October 2025, when the same judge released the company from it. In the same case the production of some twenty million conversations, in anonymised form, was later ordered.
This is not an argument against one supplier, and it matters not to read it as one: any company, in any country, is subject to the orders of that country’s courts, and anybody would have done the same because there is no alternative. It is an argument about what exactly the line you read in the documentation is worth.
That line describes what the supplier does, not what can be required of it. If your compliance rests on “they keep it thirty days”, it rests on a sentence that somebody else’s dispute, in a legal system that is not yours, can suspend without telling you. If it rests on “the document never left”, there is nothing to suspend.
Three: history and conversation memory
The part people forget because it is convenient. An assistant’s history stays in the account, and with it the document you pasted three months ago. Memory features that carry information across conversations go one step further: they extract facts from the text and keep them as a profile.
From there it takes very little for a client’s data to resurface in an answer given to another user of the same account. That is not a breach: it is the product working as designed.
A fourth place: extensions and add-ins
To the three above add one that appears in none of the main supplier’s documentation, because it is not theirs. An assistant inside the browser, an add-in for the word processor, an extension that “summarises this page”, a plugin in the practice software: each of these is a second supplier, with its own terms, its own retention and its own country.
It is the route by which documents leave an office without anybody ever opening a chat. An extension with permission to read page content also reads the practice system open in another tab; an add-in installed to fix grammar sees the text of the filing as it is being written. None of these steps looks like “pasting a document”, which is why no written rule against pasting stops them.
The only way to know is to look: the list of extensions installed in the office’s browsers, and the list of add-ins in the word processor. It is the first item on the list of permitted tools, and it is almost always the most surprising one.
Where to look, and what to write down
Three documents, in this order: the terms of the version you are actually using, not the one that appears first on the website; the data retention page, where the number of days lives; and the processing agreement, which has to exist when the data belongs to third parties (Article 28 GDPRopens in a new tab).
If one of the three answers is missing, so is the corresponding line in your record, and that is the thing an audit sees.
The question that makes all of this shorter
All three answers stop mattering the moment the text that leaves carries no personal data. A model can remember forever that [FULLNAME_1] entered into a contract with [ORG_1]: that is a sentence about nobody.
That is the point of substituting before sending, on the machine where the work happens: it is not about trusting the supplier less, it is about not having to ask them anything. The difference between local and cloud is in its own piece; what Omissia does, step by step, is in How it works.