BLOG
Use ChatGPT to review a document only when the account, company policy, and purpose allow you to send that material to the service. Reduce the information to what the review needs. If a file contains confidential client data and there is no clear authorization, do not paste it into a chat. Use an approved internal process or ask the person responsible for privacy and security.
Practical rule: replacing names is not permission to share a document. Confirm the destination and the permitted data before writing a prompt.
Classify the document before opening ChatGPT
Classify the file by the possible impact of disclosure: public, internal, confidential, or restricted. A public draft may be suitable for assisted editing. A signed contract, medical record, credential list, bank data, children’s data, or customer spreadsheet requires far stricter handling. The question is whether this specific account, configuration, agreement, and workflow permit this kind of data.
Check the organization’s policy, the purpose of the review, applicable obligations, account controls, and vendor terms. A label on a plan is not enough by itself. If authorization is unclear, do not upload the entire file. Use a local review process, an approved business tool, or obtain a documented decision from the responsible person. The following workflow is operational guidance, not legal approval to disclose personal data.
Assign a reviewer who can stop the process. That person should know what can be submitted, which account is approved, and where the final version will be stored. A convenient personal account should not become a shared repository for client documents.
Minimize information before asking for a review
Make a working copy. Remove fields that do not affect the requested check. Replace a person’s name with [CLIENT], email with [EMAIL], telephone with [PHONE], and address with a broad location if location matters. If an exact price is not needed to check structure, replace it with a range or placeholder.
Changing only a name may be insufficient. A combination of job title, date, city, unusual event, and amount can identify a person. Consider the entire context, including file names, comments, metadata, and attachments. Keep any mapping between placeholders and real people outside the chat under restricted access. Do not send the mapping to the model.
For a contract, start by reviewing the structure of clauses without signatures or client details. For a proposal, preserve the scope and decision logic while removing identifiable contacts and sensitive commercial conditions. If the needed review cannot be done without protected information, stop and choose an authorized environment instead of pretending the removal is complete.
Understand the feedback exception for personal accounts
OpenAI’s data-use guidance says people using individual plans can turn off “Improve the model for everyone” for new eligible conversations. It also says that if they submit feedback on a response, the entire associated conversation may be used to improve models even after opting out. Do not submit thumbs-up or thumbs-down feedback on a conversation containing sensitive material without understanding that exception.
The same guidance distinguishes business workspaces and API use: their inputs and outputs are not used for model training by default. That does not remove the need to check retention, access, contract terms, integrations, and sector rules. Training is one aspect of data handling, not the whole privacy assessment.
| Material | Decision before sending | Minimum control |
|---|---|---|
| Public or fictional | Suitable for testing format and instructions | Remove accidental client details and credentials |
| Internal without personal data | Use only if organizational policy permits | Check account, retention, access, and settings |
| Client or restricted | Do not send without an approved workflow | Prefer authorized business or local review |
Know what data controls do and do not do
OpenAI’s ChatGPT data-controls page explains that opting out of model improvement changes treatment of new eligible conversations; it does not by itself erase saved history. Temporary chats stay out of history and are not used to improve models while they remain temporary, but OpenAI may retain a copy for up to 30 days for safety. A user can also save a temporary chat, changing how it is handled. Check the current feature behavior before relying on a label.
Temporary mode is not the same as immediate deletion, anonymization, professional secrecy, or permission to submit someone else’s confidential file. Memory and personalization controls are also distinct from training controls. Review the settings for the actual account and workflow.
OpenAI’s Business privacy documentation describes its workspace defaults and sharing model. A business workspace can help administer a team, but it does not approve every document automatically. Determine who can see a shared conversation and whether connected tools or exports take data into another service.
Ask for a narrow, verifiable review
Ask for a defined check rather than “analyze everything.” For example: “In this anonymized proposal, identify ambiguous sentences, inconsistent dates, terms used without definition, and obligations without an owner. For each issue, cite the section, explain the concern, and suggest a question for the human reviewer. Do not invent clauses or change amounts.”
State the document type, intended reader, and review criteria. For a long file, split it into numbered sections while keeping enough context to understand each. The model’s output is a list of possible issues, not approval of the document or a final legal interpretation.
Do not ask the model to fill absent details by guessing. An absent date, permission, or clause should become an explicit question. Make uncertainty visible so the owner can resolve it from the original record.
Verify every suggestion against the original
Open the original file at each cited location. Confirm that the quotation is exact, that the suggestion respects its surrounding context, and that the model did not invent a duty, exception, number, or reference. Reject a finding that cannot be located and checked.
A simple review log can contain the section, issue type, evidence in the original, reviewer decision, and approved change. Do not paste the whole generated answer into the final document. If a suggestion touches law, safety, health, or finance, validate it with an appropriate source and qualified person.
Keep the anonymized working copy, model suggestions, and approved final document separate. That makes it possible to audit what was accepted and what was discarded without treating a draft as the signed or delivered version.
Turn a safe trial into a team procedure
Write a short procedure that people can follow: classify the file, minimize data, select the approved account, request a bounded check, verify each suggestion, dispose of working copies under policy, and record who approved the use. List prohibited document types and the person who can authorize an exception. Train with a fictional example, never by experimenting with a real client’s record.
For a pilot, define an owner, expected result, error conditions, and review deadline. Observe one execution before opening the process to a team. If an input is missing, route it to a person. Sample completed reviews for unnecessary data in prompts and for model text copied without verification.
Version the instructions and permitted document types. Revisit them when the account plan, vendor policy, connector access, or internal rules change. If several employees need the workflow, evaluate managed workspaces and permissions instead of sharing a personal login. Security comes from the full process, not a sentence in a prompt.
Common questions
Does opting out of training prevent every use of the conversation?
No. For individual plans, OpenAI says feedback may allow the associated conversation to be used for model improvement. Other handling, such as retention for safety, has separate rules.
Does Temporary Chat mean zero retention?
No. OpenAI says a copy may be kept for up to 30 days for safety while the conversation remains temporary. Saving it changes its status.
Is replacing a person’s name enough to anonymize the file?
No. Other details can identify someone in combination. Remove data that the review does not need and ask the privacy owner when the remaining material is sensitive.
Use assisted review only inside an approved boundary
The workable sequence is authorization, classification, minimization, a limited request, and human verification. Do not send restricted material when the destination or permission is uncertain. The tool can help a reviewer find issues, but it does not replace the person accountable for the document or the organization’s privacy decision.
