Before you paste it into an AI tool.

The important question comes before upload

Before asking what an AI tool can do with a document, ask whether the document should be there at all. A blank template, a public report and a customer spreadsheet can look equally easy to upload while carrying very different obligations and risks.

Start with purpose. What help do you need? If you need a formula, a few invented rows may be enough. If you need feedback on structure, a fictional version may work. If you need to analyse actual records, the decision is more involved than removing a name from the first column.

This is practical guidance for reducing unnecessary sharing, not a legal assessment, security certification or permission to upload confidential information. For work material, follow your organisation's approved tools and procedures. If you cannot establish permission, pause before sharing.

Distinguish the information from its appearance

Information can be sensitive because of what it says, who it concerns, or what other details it can be combined with. Replacing a person's name with “Person A” does not necessarily make a record anonymous. A role, location, unusual event and exact date may still identify them.

Think about what remains in the whole file. Hidden sheets, comments, tracked changes, filenames and document properties can contain information that is not obvious in the first view. A visually blacked-out area is not automatically a reliable redaction if the underlying text remains accessible.

The safest first exercise is often to create a new, minimal example rather than modify a sensitive original. Use fictional values with the same structure, and explain the relationships that matter. Do not copy unnecessary details simply because the upload box accepts a large file.

Check which account and workspace you are in

The brand name on the screen is not the whole data-handling arrangement. A personal account, an organisation-managed workspace and an API integration can operate under different terms and settings. Connected services can add another party to the flow.

Before a work task, confirm the account, workspace and approved use. Do not assume a feature is authorised because it works technically, or because somebody else in the organisation uses the same app. If an administrator manages controls, their configuration may differ from the settings shown in a consumer tutorial.

Make the question specific when asking for guidance: “May I use this approved workspace to summarise this type of document, and are there any fields I must exclude?” That is easier to answer than “Is AI safe?”

Training, history, memory and retention are different

These terms are often compressed into a single idea of privacy, but they describe different things. A training control concerns whether content contributes to model improvement. History concerns what you can return to in the interface. Memory concerns information used to personalise later interactions. Retention concerns how long data is kept under the applicable arrangement.

As one concrete example, OpenAI's data-controls guidance says turning off “Improve the model for everyone” prevents new conversations being used for training, while they can still appear in history. That control is not the same as deleting a conversation.

Its memory guidance describes separate sources and controls for remembered information. Do not assume deleting a chat removes everything remembered from it; follow the current removal instructions for your account.

These examples are about ChatGPT, not universal statements about every AI service. Controls can vary by plan, region and workspace, and their wording can change. Check the current documentation and the settings actually available to you.

Temporary does not mean no data handling

OpenAI's Temporary Chat guidance says these conversations are not used to improve its models and that a copy may be kept for up to 30 days for safety purposes. It also notes that information sent to a third party through an action is governed by that recipient's policy.

The practical lesson is broader than any one setting: do not turn a feature label into a guarantee it does not make. “Temporary,” “private” and “not used for training” answer different questions. None, by itself, settles whether you are permitted to share someone else's information.

If a workflow sends material through several tools, think about each destination. Where does the file go? Which service processes it? What can the connected tool do? Who can access the result? If you cannot describe that path, simplify the task or ask for help before using real sensitive material.

Build a minimum-useful example

Suppose you need a spreadsheet formula that assigns a follow-up category based on an invoice date and payment status. The assistant probably needs the column names, example formats, the rule and a few invented cases. It does not need customer names, bank details or the full ledger.

Write the rule separately: “If the due date is before the reference date and payment status is unpaid, return overdue.” Include an edge case such as an empty date. Use an explicit reference date so the answer is reproducible. The structure carries the problem; real identities do not improve the formula.

After obtaining an answer, test it locally on sample rows. Do not assume that a formula working on invented data is ready for every exception in the real workbook. Data minimisation reduces exposure; it does not replace validation.

Keep a short preflight checklist

Ask four questions: Do I have permission to share this? Is this the right account and tool? Does the task need all this information? Do I understand where the input and output will go?

If one answer is unclear, the next step is to resolve it, not to paste the file and hope the settings are protective. Often you can still make progress with a fictional example, a public source, or a description of the problem.

If you share something accidentally, follow the relevant reporting process promptly. Do not assume deleting the visible chat resolves every consequence. Explain what was shared, where and when, so the responsible person can assess the situation. A calm, specific report is more useful than either hiding the mistake or making unsupported assurances about it.