Most AI assistants are cloud services: the text you type or upload is sent to a server, processed there, and often kept for a while. For a meme or a recipe that is fine. For a contract, a diagnosis, a tax return, or a client file, it deserves five questions first. This checklist works for any tool, not just ours.
Open the privacy policy and find the sentence that says where processing happens. "On your device" means the model runs on the phone or computer you are holding. Anything else means a server, and the question becomes who operates it and in which country. If the policy does not say, assume a server.
Cloud services commonly keep prompts and uploads for days to years, for abuse monitoring, support, or model improvement. Look for the retention period and for a way to delete history. "We may retain" with no period is the answer to worry about. A tool that never receives the text has nothing to retain.
Three groups can usually read stored text: the provider's staff under certain conditions, contractors who review samples, and anyone who obtains access through a breach or a legal order. Check for "human review" and "used to improve our models". Business or enterprise plans often exclude training but still store the text.
Some apps advertise local features but still send telemetry, error reports, or "anonymized" snippets. The test is whether the app can function with networking off. Put the phone in airplane mode and try the feature. If it works, nothing had to leave. If it does not, something did.
Some documents come with obligations attached: health records, client matters under privilege, financial data under a confidentiality clause, personal data of other people. For those, uploading to a third party may require consent or a contract with the provider, and a "we don't train on your data" promise is not the same thing as permission.
| Question | Good answer | Bad answer |
|---|---|---|
| Where does the text go? | Stays on the device | A server, or the policy does not say |
| How long is it kept? | Never received, or deleted immediately | "May retain" with no period |
| Who can read it? | Nobody but you | Staff, reviewers, model training |
| Does it travel while you work? | Works in airplane mode | Needs a connection |
| What is the document, legally? | No third party involved | Needs consent or a contract you do not have |
On-device AI answers the first four questions by construction: the model runs on the phone, so there is no server, no retention, and no reviewer. Modern iPhones (iPhone 15 Pro and later) ship with on-device language models that handle summarizing, extracting, translating, and answering questions about a document surprisingly well, with the trade-off that they know less about the world than a large cloud model. For documents, that trade is usually the right one.
No. It hides the activity on your device; the text still goes to the provider's server.
That covers one use of the text. It can still be stored, reviewed for safety, and disclosed under a legal order. Not training is not the same as not receiving.
Safer, with contractual limits, but still a third party holding the text. Whether that is acceptable depends on the obligation attached to the document.
The model runs on the iPhone, nothing is uploaded, the app blocks outbound traffic by design, prompts and documents are never logged, and everything can be deleted in one tap. It works in airplane mode.
Quietra is an on-device AI workspace for sensitive documents on iPhone 15 Pro and later: summary, Q&A, key facts, translation into 13 languages, and version compare, with outbound traffic blocked by design. Free to start, no account.
Get Quietra on the App Store