Human review required before approval: verify every statement about training, retention, deletion, human review, connected apps, workspace controls, and account-specific settings against the current official privacy notice and the exact product surface. Confirm that this guide does not promise confidentiality, deletion, or compliance beyond the vendor’s published terms. This is a review checklist, not legal or security advice.
Before uploading a document to an AI tool, ask what the file contains, who can access it, how it may be used, and what happens after the chat ends. A setting that controls model improvement may not control storage, safety review, connected applications, or an organization’s retention policy. The safest decision begins with classifying the material and checking the current official policy for the account and feature being used—not relying on a generic “private” label.
Our editorial judgment: the most useful privacy comparison is not a single provider score. Readers should decide by data class, product surface, settings they can verify, and whether an accountable owner can approve the workflow. If one material uncertainty remains, the appropriate choice is to withhold that data until it is resolved.
1. Classify the material before choosing a tool
Write down whether the input contains public information, internal business material, personal data, credentials, customer records, regulated information, or another restricted category. Identify the data owner and whether the task truly requires the original material. If a summary, redacted excerpt, or synthetic sample can answer the question, use that instead. Do not paste secrets, credentials, or sensitive records into a consumer tool merely to test whether a control exists.
For team use, follow the organization’s approved-tool list and data-classification rules. A personal account, a business workspace, an API product, a connected app, and an embedded third-party assistant can have different terms and controls even when they use a familiar brand name.
2. Check the exact product surface and account
Open the provider’s official privacy documentation from the product you will use. Confirm whether the guidance applies to consumer accounts, a business workspace, API usage, a browser feature, or a connected integration. Check the logged-in account, plan, region, administrator settings, and feature rollout. A policy for one surface should not be copied over to another without evidence.
For example, OpenAI publishes a consumer-services data-usage FAQ and account controls; Anthropic’s help center distinguishes consumer conversations and work accounts; Google’s Gemini Apps Privacy Hub describes activity settings and connected apps; and Perplexity’s plan guidance distinguishes personal settings from organization-level controls. These pages are starting points for review, not a substitute for the terms that apply to a particular user or contract.
3. Separate training controls from storage, access, and deletion
Check each question separately: whether prompts or files may be used to improve models; whether a history or activity setting changes that use; how long conversations and uploads are retained; who may review content for safety or service operation; whether deleting a chat deletes associated files or only removes them from the user interface; and whether backup, legal, or security retention exceptions apply. Avoid treating one toggle as an answer to every data-handling question.
Record the exact control name, where it appears, the account type used, and what the vendor’s current documentation says it changes. If the documentation does not clearly answer a material question, do not infer the answer. Ask the organization’s administrator or the vendor’s official support channel before using restricted data.
4. Inspect integrations, shared links, and permissions
List every connected service, workspace, browser, file source, extension, or third-party tool that can receive the prompt or act on its contents. Review the scopes and permissions granted, who can see shared conversations or generated files, whether links are public or restricted, and how access can be revoked. Disconnect integrations that are not needed for the task. For a work account, involve the administrator when a feature can reach organizational files or identity systems.
Test with a harmless sample first. Confirm which account is active, whether connected-app context appears in the response, and whether shared outputs inherit the expected access restrictions. Do not use a test upload to discover a setting with real customer or employee data.
5. Make a go / no-go decision and retain evidence
| Decision check | Proceed only when | Stop when |
|---|---|---|
| Purpose | The task needs only the minimum necessary information. | A lower-sensitivity substitute has not been considered. |
| Policy | The official policy for the exact product surface answers the organization’s material questions. | Retention, training use, or access remains unclear for the data class. |
| Controls | Required settings and permissions are visible and can be confirmed by the right owner. | A user cannot verify the relevant settings or revoke an unneeded connection. |
| Approval | The data owner and administrator approve use under internal rules. | The material is restricted or regulated and no authorized reviewer has approved the workflow. |
Keep a dated record of the official pages checked, account and feature, relevant settings, data category, owner approval, and review date. Re-check after a policy update, product change, plan change, new connector, or change in the sensitivity of the task. A record of a setting is useful evidence of your review; it is not a certification that the tool is safe for every data type.
Who this guide is for—and its limits
Best fit: individuals and teams screening an AI tool before uploading work documents, customer materials, research, or personal information. It is particularly useful as a handoff checklist between the user, data owner, and workspace administrator.
Not fit for: anyone seeking a guarantee that a specific vendor is compliant, that a deletion request removes every copy, or that a model-training toggle prevents all processing. Those conclusions require review of current contracts, policies, settings, and applicable organizational or legal requirements.
Limitations: privacy controls differ by product, region, workspace, account, and feature; documentation and settings can change. A privacy page may not describe every processor or retention exception relevant to a negotiated contract. Escalate unclear or regulated use to a qualified privacy, legal, or security professional.
For adjacent workflows, see our AI tool selection guide, evidence documentation guide, knowledge-work tools shortlist, and ChatGPT review.
Evidence disclosure: this draft summarizes linked first-party privacy and help pages into a decision process. It does not make a comparative privacy ranking or provide legal advice. Next review: December 1, 2026.
Official sources: OpenAI consumer data usage FAQ; Anthropic consumer-account data guidance; Google Gemini Apps Privacy Hub; Perplexity official plan and privacy guidance.