What Not to Paste Into an AI Chatbot: A Privacy Checklist

Quick answer

Before sending an AI prompt, remove passwords, identification numbers, financial details, private information about people, and confidential school or workplace material. Replace necessary details with placeholders such as [PERSON_A] or [PROJECT].

Use this rule: If the chatbot can complete the task without a detail, leave it out. If the task requires sensitive information, stop and check whether your organization permits that information in that specific tool. Privacy settings are an additional precaution—not permission to share.

1. Recognize personal information and confidential material

You usually do not need to give a chatbot the original document to get useful help. For email editing, for example, start with the purpose, audience and desired tone.

Before copying anything, distinguish four terms:

Personal information: Details about an identifiable person. Identification can be direct, through a name, or indirect, through a combination of details such as location and occupation.
Confidential material: Information you are not authorized to share outside its intended setting. For this checklist, that includes unpublished business plans, internal documents and restricted classroom material—even when they contain no names.
Redaction: Removing sensitive content from the copy you intend to share, rather than merely asking the recipient to ignore it.
Data retention: How long a service keeps information. Retention, model-training preferences and saved conversation history are separate questions to check in the provider’s current documentation.

Replacing a name does not necessarily make a document anonymous. The UK Information Commissioner’s Office distinguishes anonymisation from replacing identifiers while a person remains identifiable through additional information. Treat placeholders as a risk-reduction technique, not proof of anonymity. (ICO guidance on anonymisation)

2. Use this “do not paste” checklist

For an ordinary consumer chatbot, use the following as a conservative screening checklist. It is a practical precaution, not a claim that every service handles every category identically.

Information to keep out	Examples	A lower-risk way to ask
Login and access secrets	Passwords, one-time codes, recovery codes, API keys	Describe the error without including the credential
Identity documents and numbers	Passport, Aadhaar, Social Security or student identification numbers	Use `[ID_REMOVED]`, or omit the identifier entirely
Payment and account details	Card numbers, bank account details, payment screenshots	Describe the formatting or calculation problem using fictional values
Private information about people	Personal addresses, health details, private messages, individual grades	Ask a general question or create an unrelated fictional scenario
Restricted work or school material	Customer lists, staff records, unpublished reports, confidential research	Request a generic template or use an explicitly approved workflow
Information that grants access	Private meeting links, password-reset links, access tokens	Replace the whole link with `[PRIVATE_LINK]`

Also check information about other people. Having access to a classmate’s message, a customer’s email or a colleague’s record is not a reason to include it in a prompt.

A useful alternative is to describe the task without supplying the underlying record:

“Give me a polite template for asking someone to confirm a delivery date. Include a greeting, one clear request and a closing.”

That request needs no customer name, order number or private email thread.

3. Replace sensitive details before opening the chat

Try this workflow in a local document or another environment your school or workplace approves.

Define the task. Write down what you need: “Shorten this message while keeping a friendly tone.”
Create a separate working copy. Keep the original unchanged.
Keep only the necessary excerpt. For a sentence-level edit, do not include an entire conversation.
Remove unnecessary details completely. A signature, account number or home address usually adds nothing to a writing task.
Replace necessary references consistently. Use [PERSON_A], [PERSON_B], [ORGANIZATION] and [PROJECT]. Different people should have different placeholders.
Generalize identifying context. If the task does not require an exact date or precise location, use [DATE] or a broader description.
Review the complete prompt. Read your instructions as well as the excerpt; either can accidentally contain private details.

Do not send the replacement key—the list connecting placeholders to real names—to the chatbot. Restore necessary details locally after reviewing the answer.

If removing details makes the request unclear, add task context, not personal history. For example:

“The recipient is a colleague. I want a respectful but direct request. Do not add a reason, promise or deadline.”

This keeps the writing instructions useful without restoring the information you removed.

4. Worked example: edit an email without sharing its private context

Everything in this example is fictional. The rewritten output below is illustrative, not a reported chatbot test.

Suppose this draft contains more detail than the writing task needs:

Hi Mira,
Please move my volunteer shift at Cedar Lantern Centre from 14 May to 21 May. My volunteer number is V-4821. I have a private family appointment that day. Please reply to my personal email address.
Thanks, Dev

The task is to make the request clearer and more polite—not to analyze the appointment or volunteer record.

Prepare the prompt locally:

Rewrite the following email in a polite, concise tone. Keep the placeholders unchanged. Do not invent explanations or commitments.

Hi [COORDINATOR],
Could I move my volunteer shift from [ORIGINAL_DATE] to [REQUESTED_DATE]? Please let me know whether that would be possible.
Thanks,
[SENDER]

An acceptable illustrative rewrite would be:

Hi [COORDINATOR],
Would it be possible to move my volunteer shift from [ORIGINAL_DATE] to [REQUESTED_DATE]? Thank you for considering my request.
Best,
[SENDER]

Notice what changed before submission:

Names became distinct placeholders.
The centre’s name and volunteer number disappeared.
The private explanation was omitted because it was unnecessary.
The request remained understandable.

Before sending the final email, restore the correct names and dates outside the chatbot. Check that every placeholder has been replaced and that the wording still asks for permission rather than claiming the change is confirmed.

Try it yourself: Which removed detail would improve the grammar or tone? If none would, there is no reason to put it back into the prompt.

5. Check privacy controls—but understand their limits

Controls and policies vary by service, account and workspace. Check current provider documentation before sending information. These examples are not hands-on testing or a ranking of services.

ChatGPT

Check the current privacy and data controls for the exact ChatGPT account and workspace you use. Controls and available features can differ between personal and organization-managed accounts. Do not assume a control exists or has the same effect everywhere.

Check model-training preferences, saved history, memory, retention and connected services separately. Changing a training preference should not be treated as proof that existing content was deleted. If you cannot confirm how the information will be handled, use a fictional example or your organization’s approved workflow.

Gemini

Google says Gemini temporary chats are not used to train its AI models and are retained with the account for 72 hours. With Keep Activity off, future chats are not used to improve its AI models unless you submit feedback. Safety-related processing, including human review, can still occur. (Gemini Apps Privacy Hub)

Questions to check in any tool

Before submitting information, ask:

Am I using a personal account or an organization-managed account?
What does the training control actually change?
How long is content retained?
Does submitting feedback change how the conversation is used?
Will an extension, connected app or third-party service receive information?

An extension, connected app or external service may have its own data-handling policy. Check the intended recipient and its retention terms before sending information. Do not assume a chatbot’s settings automatically apply to another recipient.

6. Watch for incomplete redaction and hidden information

The easiest mistake is to remove obvious names while leaving a recognizable story. A description such as “the only person in this small team with this unusual role” may still point to someone. Removing direct identifiers alone is not a guarantee of anonymity. (ICO guidance on anonymisation)

Another mistake is uploading an entire file after checking only its visible text. The ICO notes that electronic documents can contain hidden information, including author metadata, earlier draft details and underlying spreadsheet data. (ICO guidance on handling electronic documents)

Use these practical checks:

For email: Remove signatures, recipient lists, quoted replies and attachments unless needed.
For documents: Prefer a short, cleaned text excerpt for a writing task. If a file is necessary, follow your organization’s document-inspection procedure.
For screenshots: Inspect the whole frame, including account names, browser tabs and notifications.
For redacted files: Do not treat an added black shape as sufficient proof that underlying content was removed. Use an approved redaction method and verify the exported copy.
For earlier chats: Start with a clean task rather than adding sensitive context to an existing conversation.

Do not upload the unedited original and ask the chatbot to remove private information first. That puts the information into the service before the removal happens.

If checking the material is beyond your confidence or permissions, use a fictional example instead.

7. Follow school and workplace rules before pressing Send

Keep two decisions separate: what a service’s settings allow and what you are authorized to share.

A training opt-out or temporary chat does not override a school’s assessment rules, a workplace confidentiality policy or instructions governing a particular document. Ask the responsible teacher, manager or IT contact which tool, account and information types are approved.

Use this final check:

I am permitted to use AI for this task.
This account and tool are approved for the material.
I removed secrets, identifiers and unnecessary personal details.
I checked files or screenshots, not just the prompt text.
The remaining context does not obviously identify someone.
I understand the relevant retention and training controls.
I will review the answer and restore details locally.

If you already submitted something sensitive, stop adding information. Use the service’s deletion controls and follow your organization’s reporting process. Do not assume deleting a chat means immediate removal from every system; retention exceptions can apply.

The safest useful prompt is usually the smallest one that explains the task.