What Not to Share With AI Tools: A Practical Data Privacy Checklist

A client sends you a document and asks for a summary. You have meeting notes to organize, a spreadsheet to analyze, or an email thread to turn into an action plan.

Uploading the material to an AI tool may seem like the obvious next step. Before you do, check what the material contains, whether you are authorized to use it that way, and how the selected service handles it.

Knowing what not to share with AI helps you use these tools without including information the task does not need. The aim is to preserve useful context while keeping credentials, private records, and confidential details out of inappropriate workflows.

This guide provides a practical checklist for reviewing client files, transcripts, business data, screenshots, and internal documents before you submit them.

Quick Answer: What Not to Share With AI

Keep passwords, access tokens, recovery codes, and private keys out of general-purpose AI prompts. Do not upload private personal records, confidential client material, or restricted business information without an authorized purpose and an appropriate, approved environment.

Some organizations use AI for sensitive work under specific agreements and controls. That does not mean the same material belongs in a personal chat account or an unreviewed third-party application.

Information	Examples to exclude or review carefully	A more limited alternative
Credentials and access secrets	Passwords, API keys, authentication tokens, recovery codes, and private keys	Use placeholders such as [API_KEY] when asking for instructions.
Identity and account details	Social Security numbers, passport scans, driver’s license numbers, payment card details, and bank account identifiers	Remove unnecessary identifiers or use a fictional example to explain the task.
Sensitive personal records	Health information, employee evaluations, payroll details, student records, and private family information	Use an authorized environment when real records are necessary; otherwise ask a general question.
Confidential client and business material	Unpublished strategies, pricing agreements, financial reports, proprietary code, and internal plans	Use approved excerpts, public material, or a generic task description.
Private conversations and recordings	Meeting transcripts, support conversations, email threads, and recordings containing other people’s information	Prepare a permitted summary containing only the decisions or tasks needed.
Bulk customer or employee data	Mailing lists, customer exports, personnel spreadsheets, transaction records, and database extracts	Use fictional test rows for process questions or approved, minimized data for actual analysis.
Information hidden inside files	Comments, tracked changes, hidden spreadsheet tabs, embedded content, metadata, and identifying filenames	Inspect a separate working copy or create a clean excerpt containing only permitted content.

These alternatives reduce unnecessary disclosure. They do not automatically authorize a restricted task or make every dataset anonymous.

Why AI Privacy Involves More Than Model Training

Sending information to an AI service is a data-handling decision. You need to consider the material, the purpose, the people or services that may access it, and how long it may remain available.

The NIST Generative AI Profile identifies privacy risks involving leakage, unauthorized disclosure, and the identification of people from sensitive information.

A practical starting point is to share only what the task requires. The FTC’s Start with Security guidance similarly emphasizes limiting unnecessary personal information and controlling access.

No Training Does Not Mean No Storage

A policy about model training answers one question: whether your content is used to train or improve models. It does not, by itself, answer every question about storage, access, or deletion.

For example, OpenAI’s API data controls documentation states that API data is not used for training by default unless the customer opts in. It separately documents retention for monitoring logs and stored application features.

Those API rules do not establish the policies of a consumer chat account or a third-party application built around the API. Check the service and account you actually use.

Replacing Names Does Not Guarantee Anonymity

Changing a person’s name to “Employee A” may reduce direct identification while leaving other clues intact. A company name, unique job title, exact location, and unusual event could still reveal who the person is.

Pseudonymization replaces identifying details with labels or codes. Anonymization requires addressing whether people remain identifiable from the information and available context. The ICO’s technical guidance on pseudonymization explains this distinction.

Keep any mapping between labels and real identities outside the AI input unless its use is specifically necessary and authorized. Review indirect clues as well as names.

Check the Exact Tool, Account, and Features

Evaluate the environment you intend to use. A familiar brand name, paid subscription, or “private” label does not settle whether that particular workflow is appropriate.

Area to check	Question to answer
Authorized use	Do the relevant client instructions, workplace policies, and agreements permit this task in this service?
Training policy	How does the policy apply to this account and its current settings?
Retention and deletion	What happens to prompts, outputs, uploads, and saved project material? What exceptions apply?
Access and sharing	Who may access the material, including relevant administrators, service personnel, or recipients of shared content?
Connected services	What additional services receive data, and what permissions have you granted?
Task-specific restrictions	Are there requirements concerning record types, processing locations, or approved features?

Use current official documentation and the agreement applicable to your account. Ask the appropriate project owner or administrator when a material requirement is unclear.

For general selection criteria, see how to choose the right AI tool for your workflow. You can also record privacy requirements in the AI tool comparison scorecard.

Connected tools deserve separate attention. Asking an assistant to use one document does not necessarily restrict the permissions of an account connection. Review its actual access scope and the handling policies of additional services.

A Practical Workflow Before You Upload

Use this process whenever you are preparing material that may contain personal, client, or internal business information.

1. Define the Task Before Choosing the Input

Write down the specific output you need.

“Draft a polite follow-up email” may require a few approved facts. “Explain how to calculate a monthly total” may require a column structure and fictional rows. Neither task necessarily requires an entire email archive or financial workbook.

Ask what information would change the answer. Include useful, permitted context and exclude unrelated details.

The guide on giving AI better context and constraints can help you define those inputs.

2. Check Your Authority to Use the Material

Having access to a file does not automatically authorize sending it to another service.

Review relevant client instructions, organizational policies, and agreements. Where approval is required, confirm that it covers the intended task, information, tool, and account.

Material you own and have already approved for public use may be straightforward to assess. Private client records or internal documents may need a different process.

If a required permission is missing, use a generic task description or fictional example while the appropriate person resolves the question.

3. Choose the Minimum Necessary Input

Select the smallest input that can support a useful result:

• A public or approved excerpt.
• A short summary prepared in an authorized workspace.
• A cleaned copy containing only the relevant material.
• A fictional example for learning a method or testing a format.
• Necessary real data within an environment approved for that use.

Use fictional data to explain a process, not to produce findings about a real business. If you need actual conclusions, use appropriate real data through an authorized workflow.

Aggregation also needs review. A summary of a very small group or a distinctive transaction may still reveal private information.

4. Prepare a Separate Working Copy

Keep the original in its authorized location. Prepare the input copy before sending anything to the AI service.

Remove unnecessary names, contact information, account identifiers, signatures, personal remarks, and unrelated sections. Replace details with consistent labels when those labels are useful.

Check combinations of information that could reveal a person or organization. Retain exact dates, locations, or amounts only when they are needed and permitted.

Do not upload a restricted original to an unapproved AI tool just to ask it to remove the restricted information.

5. Inspect Hidden Content and the Final File

Review what the file contains beyond its visible main text.

Microsoft’s document inspection guidance describes comments, revision information, document properties, and hidden spreadsheet content. Work on a copy; removing hidden workbook content can affect calculations.

For PDFs, covering text with a shape is not a substitute for applied redaction. Adobe distinguishes redaction from sanitization: visible content and hidden content require appropriate removal.

Save and reopen the prepared file. Review its contents and filename. For screenshots, inspect the actual exported image for unrelated tabs, notifications, account details, or visible background information.

When a plain excerpt will do, a new document containing only that excerpt can be easier to inspect than a complex original.

6. Submit Only the Reviewed Input

Use the correct account and workspace. Confirm that the selected features and connected services match the reviewed workflow.

Submit the prepared excerpt or file, rather than adding the original for extra context. Keep later messages within the same information boundaries.

A request to “keep this confidential” is an instruction about the task. It does not change the service’s data-handling terms or settings.

7. Review the Output and Its Destination

Check the result before copying it into an email, shared document, presentation, or public page.

Look for unnecessary identifiers, confidential facts, or invented details about real people. Confirm that the intended recipients should receive the information.

Handle saved outputs and uploads according to the applicable retention process. Remove unnecessary copies through documented controls while preserving required business records.

For recurring workflows, build these checks into the review points described in human-in-the-loop AI automation.

Copy-Paste AI Data Privacy Checklist

Use this checklist before submitting prompts, attachments, screenshots, recordings, or connected-source material.

AI DATA PRIVACY CHECKLIST

Task:
Selected service and account:
Input I plan to submit:

[ ] I can explain why this task needs the selected information.

[ ] I have checked relevant permissions, policies, and restrictions.

[ ] The selected tool, account, and features fit those requirements.

[ ] I understand the relevant training, retention, access,
    and deletion rules.

[ ] I have excluded passwords, live keys, tokens,
    recovery codes, and other access secrets.

[ ] I have removed unnecessary personal and confidential details.

[ ] I have reviewed indirect identifying clues.

[ ] I have inspected hidden content and the final saved input.

[ ] Connected services receive only appropriate access and data.

[ ] I know who will review the output and who may receive it.

[ ] I know how to handle uploads and outputs after the task.

DECISION

- Proceed with the minimum permitted input.
- Use an environment approved for this information.
- Do not upload; use a generic or fictional example.
- Resolve an unanswered requirement before proceeding.

Required reviewer or approver, if applicable:
Unresolved questions:

Use the checklist to expose gaps. It is not a certification that information is anonymous, legally compliant, or secure.

A Prompt for Planning a Task Without Sharing Private Data

Use this prompt to identify the inputs a task needs before preparing them. Give it a general, non-identifying description rather than the raw document.

The prompt helps design the workflow. It does not approve the use of client information.

You are an AI workflow planning assistant.

TASK
Help me plan an AI task using the minimum necessary information.

GOAL
[Describe the output without private client or personal details.]

GENERAL MATERIAL DESCRIPTION
[Describe categories only.
Example: meeting notes containing project decisions,
contact details, and unrelated personal discussion.]

PERMITTED INPUTS
[List information categories already approved for this task.
If permissions are unresolved, say so.]

RESTRICTIONS
[List restrictions without including confidential content.]

OUTPUT FORMAT
Provide:
1. The minimum input categories needed and why.
2. Information that can be excluded.
3. Alternatives using public, generic, or fictional material.
4. A prompt template using placeholders.
5. Unresolved requirements for a human reviewer.
6. A checklist for reviewing the eventual output.

CONSTRAINTS
- Do not ask me to paste restricted original material.
- Do not infer identities or invent project facts.
- Do not declare legal compliance, anonymity, or security.
- Preserve unresolved permissions as unresolved.
- Do not access connected sources or take external actions.

Customize the Prompt

Describe the task at a useful but non-sensitive level. For example, “organize approved project decisions into an action list” provides a clearer goal than “analyze a document.”

State which input categories are permitted. You can request a template without supplying the information that will eventually fill it.

Review Its Suggestions

Check whether each requested input is necessary. Remove suggestions that would expand the data shared or conflict with your restrictions.

Verify permissions and tool requirements through the relevant people and official documentation. An assistant’s approval language cannot establish those facts.

Example: Preparing Client Meeting Notes

This hypothetical example illustrates the process; it does not describe a real client.

A freelancer wants to turn meeting notes into an action list. The original notes include contact details, unrelated personal comments, internal pricing discussion, and decisions about the project.

In this example, the client has approved a minimized project summary for use in the selected AI environment.

The freelancer prepares a separate summary containing the permitted decisions, uses role labels where appropriate, and excludes the unrelated material.

Approved project summary: The client will supply the reference images. The freelancer will prepare the first draft after reviewing those images. The project lead will consolidate feedback. A final deadline remains unconfirmed.

AI can organize this into tasks, owners, dependencies, and open questions. It should preserve the unconfirmed deadline rather than inventing a date.

The freelancer checks the result against the approved summary and maps role labels to actual contacts within the authorized project workspace.

If processing permission had been unresolved, the freelancer could still ask for a generic action-list template without sharing the meeting content.

Evaluate the Output Before Sharing It

Privacy review continues after generation. The result must be accurate, appropriate for its audience, and consistent with the information boundaries you established.

Review Checklist

• Identifiers: Does the output contain names, contact details, account information, or identifying combinations that are unnecessary?
• Confidentiality: Has it included restricted business or client material?
• Accuracy: Do statements match the approved input?
• Unknowns: Are missing decisions, owners, and dates still marked as unresolved?
• Purpose: Does each included detail help the intended task?
• Audience: Are the recipients authorized to receive it?
• Distribution: Are the sharing permissions and destination appropriate?

The AI output evaluation checklist provides a broader method for checking accuracy, clarity, and usefulness.

Improvement Prompt

Use this prompt to refine a proposed workflow or a draft already permitted in the selected environment. Do not use the review step to introduce additional restricted data.

You are an editor reviewing an AI task and its draft output.

TASK
Reduce unnecessary disclosure while preserving the task's purpose.

GOAL
[Insert a general description.]

PERMITTED INPUTS AND RESTRICTIONS
[List approved categories and relevant boundaries.]

DRAFT OR PROPOSED PROMPT
[Include only material permitted in this environment.]

MY REVIEW FINDINGS
[List unnecessary details, unsupported statements,
unclear boundaries, or unresolved decisions.]

RETURN
1. A revised draft or prompt.
2. A short explanation of removed or generalized details.
3. Remaining questions for human review.
4. A final accuracy and audience checklist.

CONSTRAINTS
- Do not introduce identifying or confidential details.
- Do not invent facts or resolve unknowns by guessing.
- Do not reconstruct removed identities.
- Do not claim that the result is anonymous or compliant.
- Do not take external actions.

Customize the restrictions and identify the specific problems you found. Compare the revision with the approved source, and confirm that it remains useful without unnecessary disclosure.

Common AI Data Privacy Mistakes

• Uploading everything for context: Select the relevant, permitted excerpt instead of an entire archive.
• Assuming file access equals processing permission: Check the rules governing the intended use.
• Using AI to clean an unapproved original: Prepare the permitted copy within an appropriate environment first.
• Removing only names: Review other details that may identify people or reveal confidential business information.
• Trusting visual coverage: Use appropriate removal methods and inspect the saved result.
• Relying on a plan label: Check the actual account, settings, features, and applicable terms.
• Treating a prompt as a privacy setting: Instructions do not replace service controls.
• Forgetting the output: Review where generated content will be stored and shared.

What to Do If You Already Uploaded Sensitive Information

Stop adding material and assess what was submitted. An accidental upload does not establish that the content became public, but it may require action.

1. Record the service, account, time, material categories, and sharing status in an appropriate restricted location.
2. Follow your organization’s incident process or contact the responsible client or data owner.
3. For exposed credentials, revoke or replace them through the issuing service and review relevant access.
4. Use documented controls to restrict further sharing or request deletion, while following any evidence-preservation requirements.
5. Contact the provider if retention or removal is unclear.
6. Seek qualified security or legal guidance where the information or incident requires it.

The FTC’s data breach response guide explains containment, documentation, and assessment. Applicable notification duties depend on the circumstances.

Do not rely on asking the assistant to “forget” as a substitute for documented deletion controls.

When to Keep the Task Outside a General-Purpose AI Tool

Keep the material out when its use is prohibited, required approval is missing, or you cannot establish that the environment meets the applicable requirements.

Regulated records, highly confidential work, and situations with unresolved contractual or security restrictions may need specialist review and a specifically approved process.

You can often continue useful work with a generic question, a blank template, or fictional test material. If removing details would make the task unreliable, use an authorized environment that can handle the necessary information.

Make Minimum Necessary Input Your Default

Understanding what not to share with AI starts with knowing what the task actually needs.

Define the job, check your authority, review the selected environment, and prepare the smallest useful input. Inspect the final file and evaluate the result before sharing it.

Build these decisions into your regular workflow. Clear boundaries help you preserve useful context while keeping control of personal, client, and business information.

Frequently Asked Questions

Does a paid AI plan make client documents appropriate to upload?

Payment alone does not establish that. Review the applicable account terms, settings, features, and project restrictions. Confirm any required approval for the actual information and purpose.

Can I rely on temporary or private chat modes?

Read the current rules for the exact mode. Check retention, access, and connected-service handling separately. A mode label does not grant permission to process someone else’s information.

Can I ask AI to anonymize a confidential file?

Only use a workflow already authorized to process the original. Uploading a restricted file to an unapproved service exposes it before any cleaning happens. Review the result for indirect identifying clues; AI cannot certify that it is anonymous.