“Use AI privately” is not a single switch. It is a chain of decisions about what leaves your Mac, which provider receives it, how long it is retained, and whether the output contains details you should not share.
Classify the material
Personal notes, source code, customer exports, credentials and public copy should not be treated as the same thing. Give yourself a simple three-level label: safe to share, share after redaction, and keep local.
Credentials and token-like strings are never prompt context. Replace them with placeholders. If a model needs the shape of a JSON response, give it a small synthetic example instead of a production payload.
Ask for less context
Models often produce better answers when the task is precise. A short prompt with a clear output format can be safer than dumping an entire folder and hoping the model finds the important paragraph.
Separate background from secrets. Include the constraints that affect the answer and remove names, addresses, account IDs and anything that would create harm if copied into a log or support transcript.
Make provider choice explicit
Cloud providers, local models and bring-your-own-key workflows have different trade-offs. Read the current retention and training terms for the provider you use. “No telemetry from this app” does not mean a cloud model never receives the prompt.
Keep a small boundary checklist near your workflow. Privacy is easier when it is a repeatable habit rather than a memory test at the end of a long day.