A useful ChatGPT setup begins with four separate choices: which workspace you are using, whether conversations can help improve models, what can personalise future answers, and which other services the assistant can access.
Those controls do different jobs. A privacy switch will not set your writing style, and a writing instruction will not restrict access to a connected app.
Check the account before the switches
Confirm whether you are in your personal account or an employer-managed workspace. Follow your workplace’s rules for approved tools and information.
OpenAI documents different data arrangements for personal and business services. Settings also vary by plan, platform and workspace policy, so an old screenshot or a colleague’s account may not match yours.
Before uploading a document, establish that this account is appropriate for the information. No combination of the settings below is blanket permission to upload confidential material.
Training and personalisation are separate decisions
Open Settings → Data controls and review Improve the model for everyone, where available. OpenAI describes this as controlling whether conversations contribute to improving its models.
Then open Settings → Personalization → Memory. Review the controls and remembered information available in your account. This concerns context used to personalise responses.
For example, you may want the assistant to remember that you prefer concise Australian English while declining to contribute eligible personal-account chats to model improvement. Those are compatible preferences because they address different purposes.
Turning memory off does not itself delete past chats. Archiving a chat is also different from deleting it.
Treat temporary chat as a specific mode
OpenAI’s current documentation distinguishes Personalized and Unpersonalized temporary chats. Check which option you select before sending the first message.
While a chat remains temporary, it does not create or update memories and is not used to improve models. A personalised temporary chat can still use existing context. The unpersonalised choice excludes memory, custom instructions and plugins from personalisation, subject to the service’s stated safety exceptions.
Temporary also does not mean immediate erasure: OpenAI says a copy may be retained for up to 30 days for safety purposes. Saving the conversation turns it into a regular chat governed by the applicable account settings.
That makes the useful choice about the conversation you intend to have, rather than an assumption that temporary mode removes every data-handling consideration.
Give writing preferences a clear job
Where custom instructions are available, use them for stable preferences. For example:
Use Australian spelling. State the currency beside amounts. Separate verified facts from assumptions. If a detail is missing, ask or mark it as unknown.
Try the instructions on a fictional enquiry. Check whether the output follows them. Do not place changing prices, customer records or passwords in permanent writing preferences.
A before-and-after test can be simple: ask for the same short supplier email with and without your instructions. Compare clarity and tone; do not assume more personalised means more accurate.
Review connected apps and shared access
Inspect which apps are connected and what permissions they have. The ability to read a file and the ability to send a message are different capabilities.
For a first writing experiment, start without unnecessary connections. If a task later needs a source, provide the smallest appropriate set of material rather than opening an entire mailbox.
Review shared links separately. Changing a model-improvement preference does not answer who can follow a link you have shared.
Keep a note of the account and settings used for a business trial. When a control or feature changes, you can repeat the test deliberately instead of wondering why the assistant behaves differently.