Prepare survey comments by reviewing identifying combinations, not just removing names. A rare role, small location and distinctive incident can point to one respondent even when direct identifiers are absent. When the useful meaning cannot be separated from identifying detail, withhold the comment from external AI processing.
Map the ways a person could be recognised
Review direct identifiers such as names, phone numbers and email addresses. Then examine indirect identifiers: branch, shift, tenure, unusual qualification, precise date and a memorable event.
Use an insider test. Could a colleague who already knows the workplace identify the speaker? In a small Indian branch office, “the only overnight technician transferred last month” may reveal more than a first name. Do not treat removal of names as a promise of anonymity.
Run a synthetic re-identification worksheet
This hypothetical comment is invented: “As the only night-shift technician at the [Small Town] branch, I missed the 17 August shuttle after returning from specialist training.”
Review it field by field:
- Rare role: “only night-shift technician.” Risk question: how many people fit? Proposed treatment: “shift worker.”
- Location: small-town branch. Risk question: does the location narrow the group too far? Proposed treatment: omit the branch.
- Incident date: 17 August. Risk question: could colleagues remember the event? Proposed treatment: remove the date.
- Training: specialist course. Risk question: does this identify one attendee? Proposed treatment: omit it.
- Analytical meaning: transport did not suit shift timing. Proposed paraphrase: “Transport arrangements did not match the respondent’s shift.”
Evaluate whether the paraphrase preserves the transport concern without retaining a recognisable story. If recognising the person remains easy, exclude it from the external dataset.
Check meaning after generalisation
Generalisation can erase the reason a minority group reported difficulty. Keep an internal, access-controlled decision record under your approved process, explaining which analytical distinctions were lost. Do not send the original-to-paraphrase mapping with the analysis copy.
Ask an authorised reviewer to assess both recognition risk and meaning loss. Use synthetic examples for tool evaluation; do not experiment with real respondent comments merely to see what happens.
Decide whether analysis should proceed
OpenAI’s supplied evidence says sharing is enabled by default on personal Free, Plus and Pro workspaces. Turning off “Improve the model for everyone” prevents training on new conversations, while those conversations can remain in history. That setting does not establish respondent permission or adequate anonymisation.
Stop when a comment remains recognisable or the proposed processing lacks approval. Use an approved internal method or report only an appropriately reviewed aggregate. If browsing the AI tools catalog, check access format and saved terms before buying; marketplace account, code or token offers are not proof of official subscription equivalence.
Sources and scope
Source material checked 29 October 2026; features/policies may change. ChatGPT data controls establishes the stated training-control distinction. The re-identification worksheet is practical editorial advice, not a legal determination or anonymity guarantee.