Consent for AI Chatbots: What a Chat Notice Has to Say Before Anyone Types
Open most AI chat tools and the first thing you get is an empty box and a blinking cursor. No notice worth the name. Then the user pastes a contract clause, a medical detail, or a customer list into it, because the interface invited exactly that and said nothing about where the text goes.
Consent for a chatbot is harder than consent for a cookie, and the reason is simple. A cookie notice describes data the site collects. A chat notice has to describe data the user is about to hand over voluntarily, in their own words, often without realising how sensitive the paste is.
Storage is the first question, not the last footnote
Every chat product stores something. At minimum there is a transcript tied to an account, plus abuse and safety logs that exist even when the visible history is cleared. A notice that does not separate those two stores misleads people. Deleting a conversation from the sidebar and deleting it from retention systems are different operations, on different clocks, and users deserve to hear that before they type, not after they ask support.
Retention periods need numbers. Kept for 30 days, kept for the life of the account, kept until a deletion request completes: any of those can be defended. Kept as long as necessary cannot, because necessary has no test a user or a regulator can run against it.
Training use is a separate choice, full stop
Whether prompts train future models is not a detail inside a general privacy policy. It changes what the user is agreeing to in kind, not degree: their words stop being a service record and become product material. That is why several data protection regimes treat model training as a purpose that needs its own basis, its own opt out where consent is not the basis, and a control that actually works without closing the account.
The control matters as much as the wording. An opt out buried four menus deep, labelled with a phrase like improve the service, will be read as concealment when a complaint lands. Put the toggle where the notice points to it, use the word training in the label, and honour it for past conversations where the product design allows it, or say plainly that it only applies going forward.
Files, voice, and pasted third parties
People upload documents about other people. A support agent pastes a customer's email. A teacher pastes a class list. The user clicking agree cannot consent on behalf of everyone inside the file, which is why enterprise and education deployments lean on contracts and data processing terms instead of a pop-up. For consumer tools, the honest move is a short warning at the upload button itself: only share material you have the right to share, and here is what happens to it.
Voice input adds a recording, a transcript, and often a speaker profile. If voice data can identify a person in the biometric sense, stricter consent rules in several jurisdictions attach to it. A notice written for typed chat does not automatically cover that.
What to test before launch
Run the deletion path yourself, in a test account, and time it. Check whether the transcript disappears from history, from exports, and from any admin view a workspace owner holds. Read the notice on a phone, where most first chats happen, and count how many taps it takes to find the training toggle from the notice text. If any of those checks fail, the notice is decoration.
Write the notice at the reading level of the people pasting things at midnight. Short sentences. Named stores, named periods, named controls.
Related reading
These earlier guides cover connected questions in more depth: Privacy Notice Version Control: Why Editing in Place Breaks Your Proof, Global Privacy Control in Mobile Apps: The Signal Most Teams Forget, and Data Processing Addendums for Small Vendors: The Consent Clauses That Do the Work.
Frequently asked questions
Does using a chatbot count as consenting to training?
No, not by itself. Use of a service and agreement to a distinct purpose such as model training are separate things, and regulators have treated bundled consent for training as invalid where users had no real alternative. A clear notice plus a working opt out or a separate consent is the defensible pattern.
How long should chat transcripts be kept?
Only as long as a stated purpose requires: support history, safety review, and legal holds each have their own period. The test is whether the service can name the purpose and the period for each store, rather than holding everything indefinitely by default.
Can an employer read staff chatbot conversations?
In a workspace deployment, usually yes within the terms the organisation set, which is exactly why staff need a notice at the point of use. Personal accounts and workspace accounts carry different expectations, and blurring them is a common source of complaints.
What should a deletion request actually remove?
The visible history, the stored transcript, and derived copies held for product use, subject to narrow legal retention duties that should be named. If safety logs are kept longer, the notice should say so and say why.