Is It Safe to Put Confidential Product Videos Through an AI Tool?

For most internal product videos, yes, provided you check four things before you upload: whether the vendor trains AI models on your content, whether that promise is in the contract rather than the marketing page, who the sub-processors are, and how long your files are retained. Tools that answer all four in writing are safe for internal recordings. Tools that answer none of them are not, regardless of how good the output looks.

The awkward part is that "is this safe?" is rarely the question a security review actually asks. It asks narrower ones, and most AI video tools have never published answers to them.

This guide is the checklist we get sent by security teams evaluating Vidocu, rewritten to be vendor-neutral so you can run it against anyone. Vidocu's own answers are included at the end, including the parts where the answer is a limitation rather than a feature.

Why this question comes up now

Screen recordings are unusually sensitive for a media type nobody classifies as sensitive.

A two-minute product walkthrough can contain a customer's name in a CRM record, an internal pricing tier, an unreleased feature flag, a staging URL, an employee's email in the corner of a Slack notification, and an admin panel that reveals your permission model. None of that is in the script. All of it is in the pixels.

The moment you route that file through an AI tool that generates subtitles, documentation, and voiceover, you have created a data-processing relationship that your security team would have reviewed carefully had you called it "sending customer data to a third party." Because it is called "making a tutorial," it often skips review entirely.

That is the real risk. Not that AI video tools are unusually dangerous, but that they get adopted without the review a comparable tool would get. It shows up most often in teams moving fast on internal enablement, where customer success and training content gets produced weekly by people who are not going to file a vendor review for a two-minute video.

The four questions that actually matter

1. Does the vendor train models on your content?

This is the question everyone asks and the one most often answered badly.

Watch for the difference between these three statements:

• "We don't sell your data." Almost meaningless. Training is not selling.
• "We don't use your data to train our models." Better, but it says nothing about the third-party models the vendor sends your video to.
• "We do not use customer data to train, fine-tune, or improve any generative AI model, and we configure our third-party AI sub-processors not to train on it either."

Only the third one closes the loop. Nearly every AI video tool is a wrapper around foundation models it does not own, so a promise that covers only the vendor's own models leaves the actual processing untouched.

Ask for the clause number in the terms of service. A vendor who cannot point to one is describing an intention, not an obligation.

2. Is the promise contractual or is it a marketing page?

Marketing pages change without notice. Terms of service change with notice, and enterprise agreements change only by mutual signature.

If the only place a vendor commits to not training on your data is a FAQ entry or a trust badge, that commitment can quietly disappear in a redesign. Ask where the claim lives. Then ask for a Data Processing Addendum (DPA), which is the document that makes data handling enforceable rather than aspirational.

For anything touching EU or UK personal data, also ask whether they publish Standard Contractual Clauses. A vendor that has never been asked for SCCs usually has not sold to a company with a real privacy function.

3. Who are the sub-processors, and can they change?

Your video does not stay with the vendor. It goes to a cloud host, a transcription model, a translation model, and a speech synthesis model, often from three or four different companies.

Two things to establish:

• The current list. A vendor should be able to produce it on request. If the list is treated as confidential, treat that as an answer.
• The change process. Most contracts, including ours, reserve the right to change sub-processors. What matters is whether you get notice and whether you can object. Enterprise agreements frequently add both. Self-serve terms rarely do.

Neither of these is a red flag on its own. Every SaaS product has sub-processors. Not knowing who they are is the problem.

4. What is the retention period, and can you shorten it?

"Deleted after processing" and "retained as long as necessary to provide the service" are very different postures, and both are common.

The practical questions:

• Can you delete a project and have the source file actually removed, not just hidden?
• Is there a configurable retention window, or is it whatever the vendor picked?
• Are retention controls available on your plan, or only on enterprise tiers?

That last one matters more than people expect. Retention controls are one of the most common features gated behind an enterprise contract, so a team piloting on a free tier is frequently operating under retention terms they would never accept in production.

The questions a security review adds

If your evaluation goes past a personal judgment call and into an actual review, expect these too.

Where is the data processed and stored? Region matters for GDPR, and "our cloud provider has EU regions" is not the same as "your data is processed in one."

Is data encrypted in transit and at rest? Table stakes in 2026, but ask anyway, because the answer occasionally reveals that only one of the two is true.

What is the authentication and access model? Who at the vendor can view your uploads, under what circumstances, and is that access logged? Support teams generally can, for troubleshooting. That is normal and worth knowing.

What happens with credentials for recorded flows? This one is newer and specific to the current generation of tools. AI recorders that drive a browser need to log into the app they are recording, which means you are handing over credentials to something. Ask how they are stored, whether they persist after the run, and whether the tool supports a session that a human authorizes rather than a stored password.

Is there a compliance attestation? SOC 2 is the usual ask. Be precise about what you are being told, because these are not the same thing:

Vendors in the third row are not being dishonest. Alignment is a real stage of a real process. But it is not an attestation, and a reviewer who reads it as one will be unhappy later.

The practical workaround most teams miss

You can remove most of the risk without resolving any of these questions, by changing what you record rather than what you use.

Record in a demo environment with seeded data. Not a copy of production with names swapped, but a genuine staging tenant with fictional accounts, fictional companies, and fictional revenue numbers.

This is worth doing even with a vendor you fully trust, because it solves a second problem: production data makes bad tutorials. Real customer names distract, real dashboards contain numbers that are wrong by next quarter, and a support agent's real inbox is visually chaotic in a way a clean demo tenant is not. Teams that build SOPs from screen recordings usually land on seeded demo environments eventually for content-quality reasons alone.

A short checklist for the recording itself:

1. Use a dedicated browser profile with no personal or unrelated logins.
2. Close Slack, email, and calendar notifications, or record in a viewport that excludes them.
3. Seed the account with fictional data before recording, not after.
4. Check the URL bar, which leaks tenant IDs and internal hostnames more often than anything else on screen.
5. Watch the recording once before uploading it anywhere.

That fifth step catches more than the other four combined.

What "safe enough" looks like by content type

Not every video needs the same scrutiny. Matching the level of review to the actual sensitivity keeps the process from collapsing under its own weight.

That last row is the one people miss. Most SaaS terms of service, including Vidocu's, explicitly forbid uploading data protected under special legislation such as protected health information or payment card data. If your video contains it, the question is not whether the vendor is trustworthy. You are outside what the contract permits, and no amount of vendor diligence fixes that.

How Vidocu answers its own checklist

Running our own questions against ourselves, with the limitations included.

Model training. Section 8 of our terms of service commits that we do not use customer data, user content, or output to train, fine-tune, or improve any foundation or generative AI model made generally available to other customers, and that we configure our third-party AI sub-processors to refrain from training on that data where the option is available. The second half is the part worth reading twice, since it is the clause most vendors omit.

Contractual, not marketing. It is in the terms, not a landing page. We publish a Data Processing Addendum and Standard Contractual Clauses for processor-to-processor transfers. Signed DPAs are available on enterprise plans.

Output ownership. You own the output generated from your content. We assign any rights we might have in it to you.

Sub-processors. The platform runs on sub-processors and our terms reserve the right to change them. The current list is available on request. If you need contractual notice of changes, that belongs in an enterprise agreement, not the self-serve terms.

Security posture. We maintain an information security program with administrative, physical, and technical safeguards aligned with the SOC 2 Trust Services Criteria, and customer videos are encrypted in transit and at rest. To be precise about the thing reviewers care about: SOC 2 alignment is in progress, and we do not currently have a completed SOC 2 attestation to hand you. If your review requires a Type II report today, we are not yet the right answer, and we would rather you learn that here than three weeks into a procurement cycle.

Retention. Data retention controls are available on enterprise plans. On other plans, retention follows the policy described in our privacy policy. Which controls come with which tier is listed on pricing.

Regulated data. Our terms prohibit uploading data protected under special legislation, including protected health information and payment card data.

Credentials for AI recording. Vidocu's AI Recorder drives a real browser and therefore needs credentials for the account being recorded, exactly as every tool in this category does. We do not treat that as a differentiator. What we do handle is the case where an account is protected by 2FA: the run pauses and asks for the one-time code rather than failing. Our recommendation is still to point it at a staging account with seeded data.

A short evaluation script you can reuse

Send this to any vendor, including us:

1. Do you use our uploaded content to train, fine-tune, or improve any AI model, including third-party models? Which clause says so?
2. Can you provide a signed DPA? Do you publish SCCs?
3. Who are your current sub-processors, and do we get notice before they change?
4. What is the default retention period, and can we configure it on our plan?
5. Is our data encrypted in transit and at rest, and in which region is it processed?
6. Do you have a completed SOC 2 report, or is alignment in progress?
7. Who at your company can access our uploads, and is that access logged?

Seven questions. A vendor selling to businesses should answer all of them same-day, and the ones who cannot are telling you something useful.

FAQ

Is it safe to upload internal company videos to an AI tool?

Generally yes for internal process and product content, if the vendor contractually commits to not training AI models on your uploads, publishes a DPA, and offers retention controls. The bigger practical risk is incidental data captured on screen (customer names, internal URLs, notifications) rather than the vendor itself. Recording in a seeded demo environment removes most of that risk regardless of which tool you choose.

Do AI video tools train their models on my uploads?

It varies, and the marketing language is often imprecise. Many tools promise not to train their own models while sending your video to third-party foundation models under those providers' default terms. Ask specifically whether the commitment extends to sub-processors, and ask for the clause number. Vidocu's terms cover both.

What is the difference between SOC 2 certified and SOC 2 aligned?

An attestation means an independent auditor examined the controls and issued a report: Type I at a point in time, Type II over a period. "Aligned" or "alignment in progress" means the controls were built against the SOC 2 criteria but no audit has been completed. Only an attestation comes with a report you can read, so ask which one a vendor means.

Can I use an AI screen recorder on an app behind a login?

Yes, but every tool in this category needs credentials for the account being recorded, so treat it as a credential-sharing decision rather than a recording decision. Prefer a dedicated staging account with seeded data over a real admin login, check whether credentials persist after the run, and confirm the tool can handle a 2FA prompt mid-run rather than silently failing.

What should I never upload to an AI video tool?

Anything protected under special legislation, such as protected health information or payment card data. Most vendor terms of service, including ours, prohibit it outright, so this is a contract limit rather than a judgment call. Also avoid unreleased financials, security architecture, and anything under an NDA that does not permit third-party processing.

The short version

The honest answer to "is it safe?" is that it depends far more on what is visible in your recording than on which vendor processes it. Fix the recording first by using a seeded demo environment, then hold the vendor to four things in writing: no model training including sub-processors, a signed DPA, a named sub-processor list, and retention you can configure.

If a vendor answers all four same-day and tells you plainly which compliance artifacts they do not yet have, that is usually a better signal than a trust badge.

Try Vidocu for free and see how screen recordings become documentation, or read the terms and DPA first. Either order is fine.