A useful privacy check begins with a concrete task. Writing a public workshop reminder needs different information from analysing a participant register. Before adjusting settings, write down the result you need and the minimum input needed to produce it.
The controls below were checked against provider documentation on October 4, 2026. They describe documented behaviour, not an independent audit. If your employer or school prohibits sending a type of information to an assistant, a temporary-chat setting does not change that rule.
1. Inspect what memory contains
Anthropic’s August memory update describes editable topics under Settings → Memory. Memory is enabled by default on Free, Pro and Max; organization owners control availability on Team and Enterprise. Shared memory applies to cloud Cowork, not local Cowork sessions.
Sensitive-topic memory is a separate choice. The same announcement describes excluding subjects such as health and beliefs unless the user opts in. Review the actual setting instead of assuming that the account default suits your task.
2. Distinguish memory-off from incognito
Claude’s current memory guide explains that disabling memory for one conversation still leaves it in chat history. Incognito conversations are excluded from ordinary history and memory, but organizational export and retention rules still apply.
The guide also separates deleting a chat from removing memory created from it. Inspect saved topics after deleting a conversation. In managed workspaces, organization-level availability and the member’s individual choice are separate steps.
3. Choose ChatGPT personalization before starting
OpenAI’s Temporary chat guidance distinguishes Personalized, which can use existing memory, instructions and plugins, from Unpersonalized, which does not. Neither updates memory while the chat remains temporary. Choose before sending the first message.
Temporary chats stay out of history and model improvement while temporary; saving converts one to a regular chat governed by account settings. OpenAI may retain a copy for up to 30 days. Workspace restrictions still apply, and limited safety-related context can still be used. Unpersonalized therefore does not mean every form of prior context is technically impossible.
Data sent through a GPT action follows the recipient’s privacy terms. The temporary-chat label does not impose OpenAI’s retention period on that third party.
4. Check the connected account, not just the app name
OpenAI’s account-connection instructions now use Settings → Plugins. Inspect the selected app and its connection details. Multiple accounts are available only where supported; access also depends on the plan, region, surface and workspace.
Review the provider permissions and available actions. Installing a plugin does not establish that the correct account is connected or that it has access to the intended records. If an app requires permissions for its selected services, choosing fewer services can narrow the connection; do not assume individual required permissions can be omitted.
A prompt naming the intended account helps communicate scope. It is not a replacement for workspace access controls.
5. Separate reading from sending
For your task, list the actions needed: read one excerpt, draft text, create a file, or send a message. A drafting task may not need access to an entire inbox. Before granting access, consider whether a permitted excerpt is enough.
Keep the review step explicit. Reading a draft is different from authorizing its distribution. Our guide to models, assistants and agents explains why these boundaries matter when an assistant can act.
Try it: prepare a workshop reminder with less data
Fictional teaching scenario: Mina coordinates a community workshop. Her organization allows public event information in an approved assistant, but forbids uploading participant records or health information. She needs a reminder draft and will send it herself after review.
Her working folder contains:
| Item | Contents |
|---|---|
| Public flyer | Saturday workshop, 10 a.m.–noon, Community Room B; bring a notebook |
| Participant register | Names, email addresses and attendance notes |
| Access notes | Individual health information and requested accommodations |
Deliberately flawed prompt:
Read all three files and both my work and personal inboxes. Write a personalised reminder for everyone, remember their needs, and send it. Use temporary chat so nothing is stored.
Before reading the answer, identify four problems and write a narrower prompt.
Answer key
- The allowed input is the public flyer. Neither the register nor access notes is needed for a generic reminder, and the stated organizational rule excludes them.
- Searching two inboxes adds unnecessary access. The flyer already contains the required facts.
- Remembering individual needs and sending messages exceeds a draft-only task.
- Temporary chat does not establish zero retention or authorize otherwise prohibited data sharing.
Corrected prompt:
Using only these public event details, draft a reminder of no more than 80 words: Saturday workshop, 10 a.m.–noon, Community Room B. Ask attendees to bring a notebook. Do not infer a calendar date or add promises about accessibility. Produce draft text only; do not access accounts, remember participant information or send anything.
Authored example output, not a tested model response:
A reminder that our workshop is on Saturday from 10 a.m. to noon in Community Room B. Please bring a notebook. We look forward to seeing you there.
Check the time, room and requested item against the flyer. The draft has no names, health information or invented date. Mina must still use the organization’s approved sending process and check the intended recipients herself.
Keep a short review record
For one permitted task, fill in:
- Result needed: What will the assistant produce?
- Allowed input: Which exact text or files may be shared?
- Account and access: Which workspace and services are needed?
- Memory and history: What do the selected controls do, and what remains retained?
- Actions: Draft only, or a separately authorized action?
- Verification: Who checks the result before it is used?
If any answer is unclear, resolve that point before supplying the data. For checking the resulting text, use How to Verify an AI Answer.
October 4 update: Replaced the August news framing with a task-based lesson, corrected connected-account availability wording, added organizational and safety-context limits, and added the fictional exercise. Provider documentation was rechecked; human editorial review remains unrecorded.