How private is AI video captioning? A creator's checklist

Creators sometimes need captions for material that is not yet public: a client announcement, an unreleased campaign, an internal interview or a draft containing personal information. The privacy question matters before upload, not after the post goes live.

The hard part is that 'private' can describe several different controls. A service can avoid public sharing but retain files for a long time. It can process transcription internally but rely on cloud storage and payment providers. It can require no account while using a secure link as the key to the job.

Map the data through the caption workflow

Begin with the artefacts a caption job can create. The original video may be uploaded to storage. Audio may be extracted for speech processing. The system can produce a transcript, timing data, caption instructions, whole-video preview media and a rendered output. Browser storage may hold a local draft or the identifiers needed to resume the flow.

• Source video and file metadata.
• Extracted or preview audio.
• Transcript text, word timings and caption styling choices.
• Preview files and the final rendered video.
• Job IDs, access tokens and recovery links.
• Payment, billing and support records.
• Browser-side state used for upload or job recovery.

A policy that only discusses the uploaded MP4 leaves important questions unanswered. Look for transcripts and rendered outputs as well, because they can contain the same speech or imagery in a different form.

Ask where transcription is processed

Some products send audio to a separate AI transcription API. Others run speech models inside infrastructure they control. The second model reduces the number of separate transcription services receiving the media, but it does not mean the whole product has no third parties.

For the standard preview flow on supported launch browsers, the initial preview transcription runs locally in your browser after it downloads the speech-recognition model it needs. The current privacy policy explains that model files may be delivered through Hugging Face infrastructure; this model delivery does not send your uploaded video or audio to Hugging Face.

Paid final processing is separate from the local browser preview, so not every transcription step happens locally. Separately, with your permission, Subvo uses PostHog analytics to understand website usage and improve the service. The wider service also uses providers for payments, cloud hosting, storage, processing and support where needed.

Those distinctions are important. Browser-local preview transcription, model delivery, paid final processing, analytics, payments and hosting are separate data flows. Ask which stage and provider a privacy statement describes instead of treating every part of the caption workflow as either entirely local or entirely cloud-based.

Check retention and deletion by artefact

Retention should be specific enough to act on. 'Temporary' is less useful than separate windows for the source upload, finished download, transcripts and billing records. Also check whether the clock starts at upload, payment, completion or last activity.

Subvo's current privacy policy says source videos are deleted within 7 days of job completion, the finished download page is normally available for 14 days, and generated transcripts, caption files and rendered outputs are deleted within 30 days unless a shorter operational window applies. Billing records can remain where law requires. Read the policy itself before relying on those windows because policies and operations can change.

Understand access links and sharing defaults

A service without mandatory accounts may use job tokens or recovery links instead of a dashboard login. This avoids storing another password, but anyone who obtains a valid capability link may be able to resume or view that job. Keep payment, job and recovery links out of public messages and screenshots.

Ask whether uploaded or rendered media is public by default, whether URLs are guessable, whether access expires and whether support staff can retrieve files. If the provider does not explain this clearly, do not upload confidential or regulated material simply because the page says no sign-up.

The no-sign-up caption guide explains the recovery trade-off in more detail. Account friction and media privacy overlap, but one does not prove the other.

Separate model training, product analytics and processing

Generating captions from your file is product processing. Using the file later to improve or train a general-purpose model is a different purpose. Look for an explicit statement about training, whether opt-in exists and whether derived transcripts or outputs are covered as well as source video.

Subvo's current privacy policy and terms say uploaded videos and generated caption or output files are not used to train general-purpose AI models unless the user explicitly opts in. This does not mean the service stores no data or uses no vendors; it is a narrower statement about training use.

Analytics is another separate layer. Subvo uses PostHog analytics to understand how visitors use Subvo and improve the website and services, and only enables it after the visitor accepts analytics. The privacy and cookie notices explain the available controls. For sensitive work, base the decision on current published terms and any contractual commitments, not an old review or marketing summary.

Privacy controls

A creator privacy checklist before upload

• Remove footage, filenames or metadata that the caption job does not need.
• Confirm you have permission to process every person's voice and image.
• Read the current retention and training terms.
• Check whether transcription uses another API or the provider's controlled pipeline.
• Identify the hosting, payment, email, support and analytics providers that matter to your risk.
• Keep job and recovery links private.
• Download the finished file within the availability window and store it safely.
• Use a contract-approved workflow for regulated, client-confidential or highly sensitive media.

If Subvo's current terms fit the sensitivity of your clip, you can start the upload flow without a mandatory account and preview the complete captioned video before deciding on the paid final export. If they do not, choose a workflow designed for your legal, client or security requirements.

For client work, write the decision down: which service was approved, which policy version was reviewed, what type of footage was allowed and who receives recovery links. This turns a vague privacy preference into a repeatable production rule and makes future policy changes easier to assess.

When the footage includes health, legal, employment, financial or children's data, a public self-serve privacy page may not be enough. Obtain the permissions and contractual assurances your situation requires, or keep the work inside an approved local or enterprise environment.