Before you paste a document into an AI chatbot, separate the information it needs from the information that happens to be there.
A chatbot helping you dispute an internet bill needs the charge, the problem, and the outcome you want. It usually does not need your account number, home address, payment details, or the entire email thread.
Keep passwords and access secrets out of ordinary AI chats. Remove unnecessary identifying details from documents. Use employer-approved tools for work information, within the rules your employer has set.
You can still get useful help. Often, a short summary gives the chatbot a clearer task while sharing less.
Keep these six types of information out of ordinary consumer chats
This is a practical default for general-purpose consumer chatbots. A specifically approved workplace or specialist service may have different protections and permitted uses. Check those before submitting sensitive material.
Information | Examples to leave out | What to provide instead |
|---|---|---|
Account access secrets | Passwords, verification codes, recovery codes, API keys, private keys, password-manager exports | Describe the problem; use placeholders such as |
Identity and payment details | Social Security numbers, passport scans, bank account numbers, complete card details | The relevant question, charge, or document wording without identifiers |
Detailed health or legal records | Identifiable medical reports, case files, private correspondence | A general question or a carefully selected excerpt with unnecessary details removed |
Other people’s private information | A child’s school records, a friend’s messages, a relative’s financial information | A general description that avoids identifying them |
Confidential work material | Customer lists, internal reports, contracts, restricted source code, employee records | Public information or a generic example, unless the exact use is approved |
Large, unreviewed collections | Whole inboxes, account exports, entire folders, long message histories | The smallest relevant excerpt or a summary you prepared yourself |
The last category matters because you may not know everything you are sending. A document can contain useful information on one page and sensitive information on another.
Never paste a secret just to ask whether it is secure
A chatbot does not need your actual password to explain password strength. It does not need a verification code to explain where that code belongs.
For technical troubleshooting, remove credentials from error messages and configuration files. Keep the structure that matters:
My request returns an authentication error. Here is a simplified example using
[API_KEY]instead of the real key.
Check URLs too. A reset link or private invitation link may contain an access token even when it does not look like a password.
If the task is creating or storing passwords, use the password features in a reputable password manager or browser. A chat conversation is an unnecessary place to keep login secrets.
Replace private details with the facts the task needs
Removing everything can make a prompt useless. The aim is to preserve the relevant facts while leaving out unnecessary identity, access, and background information.
These are illustrative examples of how to rewrite a request:
Task | Information to leave out | A more useful prompt |
|---|---|---|
Dispute a bill | Account number, address, payment details | “Draft a polite message disputing a $25 equipment fee after I returned the device. Ask what proof is needed.” |
Improve a résumé | Phone number, email address, street address, references’ details | “Improve these three bullet points for a junior front-end developer role. Keep the responsibilities accurate.” |
Organize a family schedule | Children’s names, school name, home address | “Child A has practice Tuesday from 5:30 to 6:30. Allow 20 minutes of travel each way. Flag scheduling conflicts.” |
Understand a medical term | The whole identifiable report | “Explain this term in plain English and suggest questions I could ask my clinician: [term].” |
Write a difficult message | A complete private conversation | “Help me ask a relative to return a borrowed item by Friday. Keep the tone friendly and direct.” |
Use consistent placeholders such as [PERSON A], [COMPANY], and [ACCOUNT NUMBER]. Keep the mapping in your own notes if you need to restore details later.
Preserve exact dates, amounts, or wording when they are necessary to answer the question. Replace them when they are incidental.
Removing a name does not guarantee anonymity. A distinctive job title, small town, exact event date, and unusual personal situation may still identify someone. Generalize those details when the task allows it.
Also consider the content itself. An unnamed client’s confidential business plan is still confidential.
Check screenshots and uploads, not just pasted text
A screenshot can reveal more than the sentence you want explained.
Before uploading one, look for:
Names, email addresses, account numbers, and profile photos.
Other browser tabs, notifications, or messages.
QR codes, barcodes, private links, and login information.
Addresses, school names, or location details.
Information belonging to someone else.
Crop to the relevant area when possible. If all you need is an explanation of an error, typing the error into a fresh note may be simpler than cleaning up the whole screenshot.
For documents, work on a copy. Remove unrelated pages, comments, and tracked changes where applicable. Check the filename too.
Be careful with visual redaction. Drawing a black rectangle over PDF text may cover it without removing the underlying text. Use a tool’s actual redaction function when necessary, save the result, and inspect the exported file. Check whether excluded text can still be selected or found through search.
Those checks are useful, but they are not a guarantee that a complicated file contains no hidden information. For a short question, making a new plain-text summary is often easier to review.
Do not upload the original document and ask the chatbot to anonymize it first. That shares the original before the cleanup happens.
Privacy settings help, but they answer different questions
When reviewing AI chatbot privacy, separate these questions:
Is the conversation saved in your history?
Can the provider use it to improve models?
How long is information retained for other purposes?
Can people review it under certain conditions?
Is information passed to connected tools or services?
Turning off one use does not necessarily turn off the others.
For example, Google’s Gemini Privacy Hub explains that temporary chats are not used to train its AI models, while some retention and processing still occur to provide the service and protect it.
That is a description of Gemini’s policy, not a rule for every chatbot. Check the product and account type you actually use.
A browser’s private or incognito window also does not change the AI provider’s data policy. It cannot make information you submit disappear from the receiving service.
Review connected apps separately. If an assistant can access files, email, or a calendar, check what access you have granted and whether the current task needs it.
A work subscription is not permission to upload everything
Before using AI with work material, confirm both the tool and the type of information are approved.
Ask your employer:
Which service and account should I use?
Which information may I submit?
Are uploads, connected apps, and external agents permitted?
What should I do if I share something by mistake?
Enterprise protections can differ from consumer accounts. Microsoft’s enterprise data protection documentation, for example, says covered Copilot prompts and responses are not used to train foundation models. It also describes retention, administrative controls, and separate considerations for web queries and agents.
Those protections do not replace your employer’s rules. A paid personal account is not automatically an approved workplace environment, and removing customer names may still leave restricted business information in a prompt.
If approval is unclear, use a generic example or ask the responsible team before uploading the material.
If you already pasted something sensitive
First, identify what you shared. A name in a draft and a working password require different responses.
Submitting information does not automatically mean it became public or was misused. Still, take action appropriate to the information involved.
If you shared a password or another reusable access secret: change the password or revoke and replace the exposed credential through the relevant service. Regenerate recovery codes if needed. If you reused the password elsewhere, change those copies too. The FTC’s account-protection guidance explains why exposed and reused passwords need attention.
If you shared work information: tell the appropriate security, IT, or privacy contact promptly. Explain which service and account you used and what type of information you submitted. Follow the organization’s response process.
If you shared identity, payment, or other personal records: use the provider’s available deletion and privacy controls. If the information could enable financial or identity misuse, contact the relevant institution through its official channel for guidance suited to what was shared.
Remove public share links if you created any. Check for separately saved files, memories, or connected-service copies where the product offers those features.
Use actual account controls and support channels. A chatbot saying “I have forgotten that” is not a reliable confirmation that stored data has been deleted. Deleting visible history may also be subject to the provider’s retention rules.
Do not paste the sensitive information into a second chatbot to ask whether the first exposure was dangerous. Describe the category of information instead.
Make a smaller prompt your default
Before sending, ask three questions:
Does the chatbot need this detail to do the task?
Am I allowed to share it?
Have I checked the actual text or file I am sending?
A useful starting instruction is:
Help me with [task] using this redacted summary. Use the placeholders as written. If information is missing, ask for the category of detail you need without requesting passwords, access codes, account numbers, or unrelated personal information.
That instruction can guide the conversation, but you still decide what to send.
Keep the original records in the appropriate place. Give the chatbot a carefully chosen excerpt, a generic example, or a summary whenever that is enough. You can get help drafting, organizing, and understanding information without handing over the whole file.
