Private AI voice notes: what on-device actually means
Every voice AI app has a privacy page, and they all sound alike: “encrypted in transit,” “we never sell your data,” “audio deleted after processing,” “not used to train models.” Reassuring words. They describe a system where your voice leaves your phone, gets processed on someone’s server, and you’re trusting a policy to protect it.
A policy is a promise. Policies change, companies get acquired, subpoenas arrive, breaches happen. The stronger position is an architecture where there’s nothing to promise about in the first place.
How to audit any voice AI app in two minutes
- Airplane-mode test. Turn off all connectivity and try the app’s AI features. If transcription or summaries stop working, your voice was going to a server. This test doesn’t lie: it’s the whole question in one toggle.
- Read the privacy policy for verbs. “We delete your audio after transcription” means we receive your audio.
- Check the pricing against the compute. Cloud AI costs the company money per minute of your audio. Heavy AI features plus a generous free tier usually means the business model is elsewhere, or will need to be.
- “On-device” for which part? Some apps transcribe on the phone, then send the text to a cloud model for the smart stuff. Transcript text is the sensitive part — names, numbers, the plan you haven’t announced. The question is where the AI runs, not just the microphone processing, and it sorts which “Apple Intelligence” apps are actually on-device from the ones that only borrow the phrase.
Run that audit and the field thins dramatically. On-device transcription exists in several good tools. On-device everything — transcription, the AI that thinks about what you said, even the voice that replies — is rare, because it only recently became possible on phone hardware at all.
Where Xtraktr stands, stated plainly
Xtraktr runs the entire loop on the iPhone itself: the speech recognition, the AI that asks you questions, the drafts it assembles, the voice that reads back to you. It passes the airplane-mode test with everything working. Your words never touch our servers. We didn’t have to write a deletion policy for your audio, because we never receive it.
The fine print, because an on-device pitch earns extra scrutiny: your notes sync between your devices through your own private iCloud, under your Apple ID, encrypted. That’s Apple’s infrastructure holding your data for you, the same place your photos and messages already live. Not our servers; we couldn’t read your notes if we wanted to. And the trade-off of on-device AI is real: a phone-sized model isn’t a data-center model. It’s built to interview you and structure your words, not to write essays from scratch. That’s the point, but you should know what you’re buying.
Why this changes what you’ll actually say
Here’s the underrated part. Privacy isn’t just about worst cases, because it also changes behavior at the moment of speaking. Into a cloud tool, some part of you performs; you pre-edit the half-formed thought, skip the client specifics, soften the honest read. Think of the post-client-call debrief, the competitive take, the idea that might be bad. Those are the thoughts that most need talking through, and exactly the ones you won’t hand to a server.
A thinking partner only works if you’ll think out loud in front of it, unedited. That’s not a feature. It’s the precondition for all the others.
