Understand where voice-assistant data is processed

List the inputs you intend to share: live speech, pasted text, a screenshot or a document excerpt. Then identify the selected provider and the device. A photo may contain names and addresses that your spoken question never mentions, so inspect the whole input before sending it.

For GPT Realtime Assistant, the public mobile workflow uses online sessions. The public description says session audio goes to OpenAI and sessions are stored on the device. That distinction does not prove that every provider or backend record has zero retention, and this guide does not replace the app’s privacy policy.

Processing is where the request is interpreted or answered. Storage is what persists afterward, for example a device session, a file you saved or service metadata. An online model can produce an answer that is later kept locally. Conversely, local speech recognition does not mean a later text-analysis request stayed local.

OpenAI’s API data-control documentation distinguishes abuse-monitoring records from application state, with behavior depending on the endpoint and eligible account controls. Zero Data Retention is not something to infer from an API key or from an app calling a service. The publisher’s actual configuration would need confirmation.

For a proposed meeting-note task, write four lines: what is captured, who participates, which service processes it and where the result will be kept. Replace real names with role labels when identity is unnecessary. Ask permission before including other people in a session.

An example safe test is a fictional project discussion, not a recording of a private customer call. If a work policy prohibits sending the material to external AI, a preference for convenience does not change that restriction. Use an approved workflow or remove the restricted material.

Microphone permission controls capture access. A speech-output toggle controls what is played aloud. Deleting a local session concerns a stored history. Account or provider deletion requests concern other records. None of those actions should be treated as a substitute for all the others.

If the app’s available controls do not answer a retention question, ask support a precise question: which data category, which processing mode and which device? Avoid sending more private content just to demonstrate what you want deleted. Account verification should follow the publisher’s support process.

Use a short sanitized excerpt rather than a complete document when it answers the question. Crop a screenshot and review what remains visible. Keep critical facts in a form you can verify without uploading an entire record. These choices reduce exposure even when a task legitimately uses cloud processing.

Do not treat an offline label from one mode or device as a blanket statement about all app activity. Subscription checks and other service functions may be separate from local inference. When the workflow is uncertain, pause the sensitive request and use synthetic material until the processing path is clear.