Pasting a document into a hosted AI chatbot submits its contents to a service. Whether that service also stores the chat, uses it for model improvement or lets people review it depends on the product, account and settings. “History off” is not a universal privacy guarantee.
The useful decision comes before submission: can this particular account handle this particular material under acceptable terms? A helpful answer afterward does not answer that question.
Four questions, not one privacy switch
Processing: What content will the service receive to answer the request? Include attachments, not just the sentence typed beside them. For example, Google’s Gemini Apps Privacy Hub lists submitted prompts and shared files among the information it collects.
Retention: How long is that content kept, and what survives deletion? Model improvement: Is it used to improve models, and which opt-out applies? Human access: Can people review it for quality, safety or feedback handling? These questions overlap, but they are not interchangeable.
The same Google documentation provides a concrete example. With Keep Activity off, future chats are not used to train its AI models unless the user submits feedback, but they are still retained with the account for 72 hours. Google also says human review can support safety even with the setting off. Chats already reviewed by service providers can be retained for up to three years and are not removed by deleting account activity. The hub separately says these settings do not control processing chats to create anonymized data to improve Google services. Those are specific policy conditions, not rules for every chatbot.
See how this works: four data questions
Check the account you actually have
For a personal Gemini account, the activity-control guide provides the route to Gemini Apps Activity and the Keep Activity setting. Read its current explanation before changing it. If you want previous activity removed as well, do not confuse turning future collection off with deleting earlier activity.
Work or school accounts are different: the same guide says the administrator controls activity settings, and users may not be able to change retention or see its duration. A missing control is a reason to ask for the applicable policy, not to assume retention is disabled. A paid personal plan is not, by itself, evidence that an organization’s agreement applies.
Make a version that does not need the original
Hypothetical example: You want a clearer structure for a confidential purchasing document. The job is organizing requirements, not evaluating its confidential terms. Instead of uploading it, write: “Create headings for a fictional purchasing checklist covering requirements, approval, delivery and acceptance. Use placeholders and do not add prices or deadlines.”
That request supplies the shape of the work without the document. If a specific sentence needs editing, create a synthetic sentence that illustrates the same writing problem. Avoid merely swapping an obvious identifier while keeping unusual figures and circumstances intact. A recognizable story is still recognizable with a label changed.
For an approved excerpt, place only the necessary text in a fresh plain-text draft and read it again before copying. This makes the intended submission inspectable. It does not make sensitive words safe or override confidentiality rules. If the task genuinely requires restricted source material, use an approved service and account or do the work without submitting it.
The finish line is a decision you can explain: the account’s processing, retention, model-improvement and review terms are known, and the exact submitted material is permitted. Reopen the setting to confirm a change saved. A chat disappearing from history verifies that interface action, not the erasure of every retained copy. If the terms remain unclear, keep the original out of the box.