Is your footage used to train AI? Reading the fine print

Is your footage used to train AI? It depends on the tool, the plan you are on, the feature you used and the settings on your account, and the only reliable answer is in the vendor's own terms. Look for three things: whether content you upload or process may be used to "improve" products or models, whether that is on by default, and where the opt-out lives. Footage that never leaves your Mac cannot be used for anyone's training, so the safest workflows are the ones that keep processing local.

This guide walks through how to read the paperwork, what the common phrases usually mean, how to switch training off where a vendor allows it, and which kinds of work avoid the question entirely.

Why the answer is rarely a simple yes or no

Software companies that offer AI features have a reason to want real-world data: models get better with more examples. They also have strong reasons not to take it carelessly, because professional customers work under NDAs, release windows and privacy law. The result is a patchwork.

• It can differ by plan. Business, team or enterprise agreements sometimes carry stronger data commitments than individual plans. Check which terms apply to the account you actually use.
• It can differ by feature. A desktop editor may process most work locally while one cloud feature, such as generation or an assistant chat, sends media or prompts to a server with its own rules.
• It can differ by storage. Files you put in a vendor's cloud storage, shared review links or a stock contribution programme may be covered by different clauses from files that stay in a local project.
• It changes over time. Terms are updated, and AI clauses are among the most frequently revised. A policy you read last year may not be the one in force now.

So the useful question is narrower: for this account, this feature and this storage location, today, may my content be used to train or improve models?

Where to look

Vendors usually spread the relevant wording over several documents. Open each one and search rather than reading top to bottom.

1. Terms of use or general terms. This is where the licence you grant over your content is defined.
2. Privacy policy. Describes personal data, which can include people who appear in your footage and the transcripts of what they said.
3. AI or generative features page. Many vendors now publish a separate FAQ or policy for AI features. It is often the clearest source.
4. Product-specific terms. Cloud storage, collaboration and stock programmes may have their own.
5. Account privacy settings. The actual switch, if there is one, lives here, not in the legal text.

Write down the date you read each document. If a project depends on the answer, save a copy or a PDF with the project paperwork.

The words that matter

Search each document for these terms and read the full sentence around every hit.

• "train", "training", "machine learning", "model". The direct mention. Note whether it says your content is used, may be used, or is not used.
• "improve our services" or "develop new features". Broad wording that can include model improvement even if the word "train" never appears. Look for a nearby sentence that narrows or excludes AI training.
• "content analysis" or "analyse your content". Covered in detail below.
• "licence", "worldwide", "royalty-free". Most services need a licence to host and process your files. Check what the licence is for: operating the service, or also developing products.
• "aggregated", "de-identified", "anonymised". Means the vendor may use data in a form that is meant to stop it pointing back to you. For video, ask yourself what de-identification of a frame would really look like.
• "opt out", "opt in", "settings". Tells you whether there is a choice and which way the default points.
• "human review". Whether staff may look at content, for example to check quality or safety.
• "third party" or "subprocessor". Whether another company's model handles requests, and whether its terms add anything.

A clause that says content "is not used to train generative AI models" is not the same as one that says content "is not used for any machine learning". Read the exact scope.

What content analysis usually means

"Content analysis" is one of the most misunderstood phrases in software terms. In general, it describes automated processing of your files by the vendor's systems. That can cover several quite different things:

• Running the feature you asked for. Transcribing a clip, detecting scenes or tagging a photo is technically analysis of your content.
• Safety and abuse checks. Scanning hosted files for illegal material or terms violations.
• Product improvement. Using patterns in content to build or tune features, which in some policies can include machine learning.

Because the phrase covers so much, it is not evidence either way on its own. Look for what the policy says the analysis is for, whether it applies to files stored in the vendor's cloud only or also to local files, and whether the improvement purpose can be switched off. If the wording leaves you unsure, ask the vendor's support or sales team in writing which category your use falls into, and keep the reply.

How to opt out where you can

Where a vendor offers a choice, the steps tend to look similar:

1. Sign in on the web and open the account, privacy or data settings page. These switches are often not inside the desktop app.
2. Look for wording such as "content analysis", "product improvement", "help improve", "use my data for training" or "machine learning".
3. Switch it off and take a screenshot with the date visible.
4. Repeat for every account. Team members, freelancers and assistants each have their own settings unless an admin enforces a policy.
5. Check team or organisation controls. On business plans an admin may be able to set this centrally, which is more reliable than asking everyone.
6. Recheck after major updates or when the vendor emails you about changed terms.

An opt-out generally applies going forward. Whether it affects data already processed is something the terms may or may not address, so read that part too, or ask.

Workflows that never expose your footage

The surest way to keep footage out of training is to keep it out of other people's systems. That does not mean avoiding AI, because many useful jobs now run on the computer itself.

• Edit and render locally. Keep the project file and media on your own drives, and treat cloud sync as a deliberate choice for specific projects rather than a default.
• Prefer on-device AI for sensitive material. Transcription, subtitles, search, face blur and noise reduction can run locally in a number of current editors. Test each feature with the network off, as described in does AI video editing upload your footage.
• Share deliberately. Send review copies through a channel whose terms you have read, and watermark or limit what you send for unreleased work.
• Keep generative cloud tools for material you could publish. If a feature has to upload, feed it only footage you would be comfortable seeing outside the project.
• Check crash reports and diagnostics. Some bundles can include file names or paths. Look for a way to review or redact them before sending.

For the wider case for local processing on films with confidential subjects, see on-device AI video editing and privacy.

Where cinemaOS fits

cinemaOS is a Mac-native editor for Apple silicon, with its AI features designed to run on the Mac. Projects are local-first .mntm files on your own drives.

J.A.R.E.D., the on-device assistant, proposes changes, such as a look you describe in a sentence, and with Always ask enabled nothing is applied until you approve it. It uses no generative credits.

Transcription, subtitle creation and local subtitle translation run on the Mac. The offline subtitles tutorial translates captions into French with Wi-Fi switched off, which is a quick way to see the local claim for yourself.

The limits are worth stating too. cinemaOS does not do lip sync, does not open Premiere Pro project files directly, and is not a full After Effects replacement. It runs on Mac only for now, with Windows planned for November 2027. And like any vendor, it should be tested rather than taken on trust: the offline test works on it the same way it works on everything else.

FAQ

If I opt out, is my old footage removed from training data?

Not necessarily. Opt-outs usually apply from the moment you switch them off. Whether earlier data is excluded or deleted depends on the vendor's terms, so read that section or ask support in writing.

Does a local desktop editor ever send footage for training?

The editing itself runs on your computer, but individual cloud features, cloud storage and some diagnostics can send content elsewhere. Check each feature with the network off and read the terms for any feature that needs a connection.

What should I tell a client who asks whether their footage trains AI?

Tell them which tools touched the footage, which features ran locally, and which training settings you checked, with dates. A short written note is usually more reassuring than a general promise, and it gives you something to update if a tool changes its terms.