There is one test that settles this, and it takes less time than reading a privacy policy: put the phone in airplane mode and dictate a sentence. If text appears, the model is on the phone. If nothing appears, the model is on a server and the app has been telling you about its policy, not its architecture.
Everything else in this post is detail around that test.
Why the phone is the hard case
A laptop can hold a large model, run it on a fan-cooled chip, and stay plugged in. A phone has a battery, no fan, and a thermal budget measured in seconds of sustained load. That is why almost every transcription app you have heard of runs the model on a server and uses the phone as a microphone with a screen.
The trade is legible: cloud models are more accurate on difficult audio and much faster to improve, because the vendor ships a new model by changing a server. On-device models are fixed at install time, limited by the device, and cannot be improved for you without an app update.
What you get for that trade is the thing the cloud cannot offer: the audio never leaves the device.
What that is worth on a phone specifically
- No connection required. Basements, planes, hospitals, factories, and buildings that block outbound audio all work the same as a desk.
- No copy on a server. A breach at the vendor contains nothing of yours, because nothing of yours was sent.
- No retention window to configure. There is no server-side history to keep, delete, or forget to delete.
- No carrier in the path. Cellular networks are a third party. So is hotel Wi-Fi, and so is the captive portal that wants your email before it lets you upload a recording of your therapist.
What Jot does on iPhone
Jot is a keyboard-first transcription app for iOS 26.0 or later, on iPhone, iPad and iPod touch. It transcribes speech on the device, inserts the result where you are typing, and keeps drafts and transcripts locally.
The App Store listing states the architecture directly: “No text is saved, transcription is done on device, corrective AI calls are de-identified, use American ZDR providers, and optional.”
The optional part is the corrective pass. When you turn it on, Jot can send a de-identified text snippet to a US language-model provider under a zero data retention agreement — no audio, no account, no identifier. Turn it off and the app makes no transcription calls at all. It is a single purchase, $49.99 once, rather than a monthly seat.
How the iPhone options compare
| App | Where the model runs | Works offline | Price model |
|---|---|---|---|
| Jot | On the iPhone | Yes | $49.99 once |
| Wispr Flow | Cloud — “transcription always occurs on the cloud” | No | $144 a year (Pro, annual) |
| Otter.ai | Cloud | No | $99.99 a year (Pro, annual) |
| Fireflies.ai | Cloud | No | $120 a year (Pro, annual) |
| tl;dv | Cloud | No | $216 a year (Pro, annual) |
The prices are the vendors' published annual plans, and they are per user. Every one of the four cloud products above costs more in its first six months than Jot costs ever.
Short answers
Does on-device transcription work in airplane mode?
Yes. That is the definition of on-device: the model runs on the phone, so airplane mode changes nothing. Test it by turning on airplane mode, then dictating a sentence.
Is Jot's transcription really on the iPhone?
Yes. Jot runs speech-to-text on the device and keeps drafts and transcripts locally. Its optional corrective AI pass sends de-identified text, never audio, to providers under zero data retention agreements.
Can a cloud transcription app be private?
It can be responsibly run, with encryption, retention limits and contractual guarantees. It cannot make the upload disappear, so the vendor's server is always part of your threat model.
The honest trade
On heavily accented speech, in a noisy room, with overlapping speakers, a frontier cloud model still wins. Jot's on-device model is good at the job most people actually have — one person, one microphone, ordinary acoustics — and it is not going to beat a server GPU at the hardest 5% of recordings.
What it does beat is the assumption that every recording has to be uploaded. Most do not, and the ones that do are usually the ones you would least want on someone else's disk.
Verdict
Run the airplane-mode test on whatever you use. If it fails, the app is a microphone with a server behind it, and its privacy claims are contract terms rather than architecture.
If it passes, ask what the network is still used for. Jot uses it only for an optional, de-identified corrective pass that you can switch off.
If you want the iPhone to do the transcription, Jot on the App Store — $49.99 once, no subscription.
Sources
Primary sources, read on the date above. Competitor prices are the vendors' published annual plans, per user.