On-device AI video editing: why privacy matters for film

On-device AI means the model runs on your own computer and your footage, audio, transcripts and prompts never leave it to be processed. For filmmakers this is not an abstract preference. An unreleased feature, a cut under NDA or an interview with a vulnerable documentary subject is exactly the material you should not be sending to a third-party server just to get subtitles or a rough assembly.

Many AI editing features today send media to the cloud for transcription, translation, voice or generation. That can be fine for a public vlog. For anything confidential, the question to ask is simple: where does the processing happen, and what is kept afterwards. This article explains why it matters, what runs locally in cinemaOS, what still uses the network, and gives you a checklist for evaluating any AI tool.

Why local processing matters in film work

Unreleased films and NDAs. Production and distribution contracts often restrict where footage may be stored or sent. A clause written before AI tools existed may still cover an upload to a transcription service. If you cannot say where a copy of your picture went, you may already be in breach.

Documentary subjects. People who speak on camera about health, immigration status, abuse or politics trust you with that material. Sending their voice and face to a service whose retention policy you have not read is a decision about their safety, not only yours.

Leaks and spoilers. Every additional place your cut lives is another place it can leak from. A transcript is a script of your film. Treat it with the same care as the picture.

Training and retention. Some services reserve the right to keep uploads or use them to improve models, sometimes by default, sometimes behind a setting. Policies change. Local processing removes the question.

Working offline. Location shoots, trains, festival hotels with poor Wi-Fi. If the AI runs on your machine, it works there too.

What "on-device" should actually mean

The term gets stretched, so be precise. A genuinely local AI feature should:

• Run the model on your computer's own processor and memory.
• Read media from your disk and write results back to your disk.
• Work with the network switched off.
• Not send the prompt you typed, the transcript it produced, or thumbnails of your frames anywhere.

A feature that runs a small model locally but sends your request to a server to "understand" it is not fully on-device. Neither is a tool that processes locally but syncs results to a cloud project by default.

A checklist for evaluating any AI tool's privacy

Use this before you put confidential material into any editor, plug-in or web service. None of it is specific to one product.

1. Turn off the network and try the feature. This is the most honest test there is. If transcription or translation fails offline, it is not local.
2. Read the privacy policy for the words "upload", "retain", "train" and "improve". Note how long data is kept and whether you can opt out.
3. Check the default settings, not only the possible ones. A setting you have to find is a setting most of your team will not change.
4. Ask what is sent for account and licensing. Almost every paid app talks to a server for that. That is normal; it should not include media or content.
5. Look at crash and diagnostic reporting. Find out whether logs can include file names, transcript text or frames, and whether reporting is optional.
6. Check collaboration features. Shared review links and cloud projects are often where media ends up on someone else's server.
7. Confirm what happens to voice data. Voice cloning and dubbing tools should ask for consent and tell you where voice models are stored.
8. Write it down for your producer. A short note saying which tools touched the footage and where processing happened is useful for any NDA or legal review.

If your concern is protecting people on screen as well as the files, our guide to blurring faces in video on a Mac covers anonymising subjects before a cut goes out.

What runs on-device in cinemaOS

cinemaOS is a Mac editing and finishing app for Apple silicon. Its AI features are built to run locally, and no project, media, transcript or prompt is uploaded for AI processing.

Transcription. Speech is transcribed on the Mac using bundled Parakeet and Whisper models, with speaker detection for up to four speakers. Press ⇧⌘T ("Transcribe with J.A.R.E.D.") or use Finisher ▸ Localize ▸ Captions.

Translation. Subtitles can be translated locally into 58 languages with Translate Local. The app labels quality honestly: Dutch, French, German and Portuguese (Portugal) are rated Good, and other languages are Draft or Rough and need human review before delivery. More on that in translating subtitles locally on a Mac.

J.A.R.E.D. The assistant (⌘J) runs a small local model on the Mac. You type a request such as "Tighten this interview to 2 minutes", and the edits land on your real timeline where ⌘Z undoes them. Destructive actions always ask first. It understands English best today; other languages are noticeably less reliable.

Voice. Dubbing uses the voices installed in macOS. Optional voice cloning only runs after you complete a voice-rights attestation, transfers timbre only, and writes a provenance sidecar file next to the result. There is no lip sync.

Analysis and finishing. Reframe subject tracking, face and object covering, restoration and upscaling models are bundled with the app and run locally. On 8 GB Macs, heavy jobs run one at a time.

You can run the offline test from the checklist yourself: switch Wi-Fi off and transcribe or translate a clip.

What still uses the network

Being honest about this matters as much as the local part. cinemaOS uses the network for:

• Your Momentum Account (sign-in, email verification).
• Licence checks. The app works offline for up to seven days between checks. Projects open and save even without a licence.
• Checkout and billing.
• Updates. The app checks for updates and never installs silently.
• Import From URL, when you choose to download a media file from a link.

None of these carry your project, media, transcripts or prompts. There is also no cloud project sharing or real-time collaboration in cinemaOS, so there is no shared server copy of your cut. Review notes stay inside the project file on your Mac.

FAQ

Is local AI as good as cloud AI?

Not always. Local models are smaller, so some tasks are less polished, which is why cinemaOS labels translation quality by language. For confidential work, a slightly rougher draft you review yourself is often the better trade.

Does on-device AI need a powerful Mac?

It needs Apple silicon and macOS 15 or later. 8 GB of memory is supported, with heavy AI jobs running one at a time.

Can I prove to a client that nothing was uploaded?

You can show the tool works with the network off, and document which tools you used. That is usually what an NDA review asks for.

Test it with your own footage

Pick a clip you would not want to upload anywhere, switch off Wi-Fi, and see what your current tools can still do. To try the same test in cinemaOS, the 60-day free trial needs no card.