A dictation tool is the one app on your Mac you talk to all day, which means it's also the one app that hears everything: the confidential number in a draft email, the client name you'd never put in writing, the password you muttered while reading it off a sticky note, the half-formed sentence you deleted. Text you type stays between you and the keyboard. Speech has to travel — from your microphone, through some pipeline, back as words on screen — and every step along that path is a place your voice could be kept, copied, or reused. Most tools describe what they do; far fewer describe what becomes of what you said. Here are the questions that tell you the difference, and why each one matters.
Is the audio kept, and for how long?
The first question is the simplest and the one most easily glossed over: after your speech becomes text, does the recording still exist? Turning audio into words doesn't require keeping the audio afterward. But a tool that retains recordings — to "improve quality," to let you replay a transcript, or with no stated reason at all — is holding the rawest version of everything you said, in a form no amount of editing the text can take back. You can fix a wrong word in the output. You can't unsay it in a file on a server.
So ask plainly: is the audio retained, where, and for how long? "Deleted immediately after transcription" is a very different posture from "retained for 30 days" or a policy that never mentions audio at all. The strongest answer is that the recording is used to produce your text and then discarded — never written to disk as a keepsake. Sageio Type is built around exactly that intention: audio exists to become your words, not to be stored. No audio retention is treated as a design goal, not a setting you have to remember to switch off.
Is your speech used to train models?
The second question is where "free" often gets expensive. Your voice, your phrasing, the specific things you dictate — all of it is valuable training data, and a tool that uses your speech to improve its models is getting something real from you in exchange for the service. That can be a fair trade if it's disclosed and optional. It's a problem when it's the default, buried in a policy, impossible to turn off.
Beyond principle, training use means your actual words may pass through human review, sit in datasets, and shape a model other people use — confidential phrasing doesn't stay yours. Look for a clear statement that your dictation is not used to train models, and be wary of the softer "we may use anonymized data," which can mean more than it sounds. Sageio Type is designed so what you dictate is yours, not raw material for training.
Where does your history live?
Everything you dictate accumulates somewhere, and where is a security decision disguised as a convenience feature. If your history — the running record of everything you've spoken into the tool — lives on a company's servers, your account becomes a single point of exposure: a breach, a subpoena, an over-broad internal query, or a bad password all reach the same archive. If it lives on your Mac, the blast radius is your Mac, which you already secure with a login, disk encryption, and physical control.
Cloud history buys you sync and a searchable archive — a genuine benefit, and for some people worth it. But it should be a choice you make knowingly, not a default you discover after the fact. Ask where the history is stored, who can read it, and whether you can delete it for good. Sageio Type is designed to keep your history local — on the machine where you dictated it — so the record of what you've said stays under the same lock as the rest of your Mac, not in an account you trust someone else to guard.
What actually leaves the device?
"On-device" and "in the cloud" get used loosely, and the honest answer for most capable dictation tools is some of both — so the useful question isn't a slogan, it's a map. What specifically leaves your Mac, to where, and in what form? Raw audio or already-processed text? A short burst for one sentence, or a continuous stream to a service the policy won't name?
A tool that leans on remote processing isn't automatically worse — remote models can be more capable — but you deserve to know the shape of what's transmitted so you can decide what you're comfortable dictating. The right posture is minimalism: send only what the task needs, be specific about what that is, and don't route your words through more hands than the work requires. If a tool can't tell you plainly what leaves your device, treat that vagueness as the answer.
What controls and permissions does it ask for?
The last question you can check yourself, today, without reading a single policy: what does the app ask your Mac for? A dictation tool has an honest, narrow need — the microphone, so it can hear you, and accessibility permission, so it can type into other apps. A tool reaching for more than its job requires is telling you something about how it thinks. macOS shows you every permission an app requests, and that list is a plain-language summary of its real intentions, no marketing in between.
Alongside permissions, look for the controls that put you in charge after the fact: can you delete your history, export it, close your account and have the data actually go away? Those controls signal a tool that expects you to stay in control of your own words. Sageio Type is built to ask only for the microphone and accessibility permissions it needs, and nothing beyond that — the narrow request is deliberate. You can read more at sageio.net/type.
The one question underneath all five
Every question above is really one question wearing different clothes: does this tool treat your voice as yours, or as an asset? You won't find the answer in a feature list, because privacy isn't a feature — it's a set of decisions a tool made before you installed it. Run any dictation tool through these five questions before you talk to it all day. The ones built to respect you will have clear answers. The ones that hope you won't ask are telling you something too.