Writing an AI Disclosure Policy for Your Company

The question that finally forces the policy is always the same, and it always arrives on a Friday: "Do we have to say this ad used an AI presenter?" Someone Slacks legal, legal says "probably, depends," the launch slips a week, and two months later the exact same question gets asked again by a different person about a different asset.

A written AI disclosure policy is not a compliance document. It's a decision cache. You answer the hard question once, in calm conditions, and then every creator on the team answers it in four seconds by looking at a table. Companies that skip this don't end up more aggressive — they end up slower and inconsistent, which is worse for the brand than either extreme.

Start from what disclosure is for

There are two separate reasons to label AI content, and confusing them produces bad policies.

Reason one: platform and regulatory compliance. Certain platforms require labels for realistic synthetic media, particularly anything depicting real people or real events. This is a rules question with a rules answer.

Reason two: audience trust. Your audience's tolerance is not the same as the legal minimum. A B2B software buyer generally does not care that your product explainer used generated B-roll. That same buyer will feel misled if a "customer testimonial" turns out to be a synthetic person. The betrayal isn't the AI — it's the implied claim that a real customer said that.

The useful test: would a reasonable viewer feel deceived if they found out? If the answer is yes, disclose regardless of what the rules require. If the answer is obviously no — an animated logo sting, a generated background texture — labeling it just trains your audience to ignore labels.

The four-tier disclosure model

Give your team a ladder, not a judgment call. This is the structure I've seen hold up across consumer and B2B teams alike.

Tier	Content type	Disclosure required	Example
0 — None	AI used as a production tool, nothing represented as real	No label	Generated B-roll, backgrounds, music beds, upscaling, captions
1 — Light	Synthetic visuals that are clearly stylized or fictional	Caption-level note or platform toggle	Stylized brand mascot, illustrated explainer
2 — Explicit	Photorealistic synthetic humans, voices, or scenarios a viewer could mistake for real	On-screen label plus platform toggle plus caption	AI presenter, cloned voice narration, generated "street interview"
3 — Prohibited	Content that implies a real endorsement, event, or result that did not happen	Do not publish	Fake testimonials, synthetic executives of other companies, fabricated results footage

Two things make this work. First, tier 0 has to genuinely exist — if every use of AI requires a label, people stop reporting honestly. Second, tier 3 has to be enforced. A policy with no "no" in it is a suggestion.

Match the tiers to platform rules

Platform requirements shift, so your policy should reference categories, not paste screenshots that go stale. As of mid-2026, the consistent pattern across the major networks:

• Most platforms provide an AI-generated content toggle at upload. Using it is cheap and reduces the chance of an automatic label being slapped on with less flattering wording.
• Requirements tighten sharply for realistic depictions of real people, elections, health, and financial outcomes.
• Ad review is stricter than organic review. An organic post that passes may be rejected as an ad.
• Some platforms detect provenance metadata automatically. Stripping metadata to avoid a label is the single worst move available to you.

Write the policy as: "Tier 2 content uses the platform AI toggle on every network where one exists, plus an on-screen label in the first three seconds." Then maintain a one-page appendix of which networks have which toggle, owned by whoever runs publishing across your nine platforms. Appendices go stale gracefully; policy bodies do not.

Label wording that doesn't hurt performance

Teams resist disclosure because they assume it costs reach. In practice the wording matters more than the presence of a label.

Wording that reads as defensive:

• "This video contains AI-generated content."
• "Disclaimer: synthetic media."

Wording that reads as confident:

• "Made with AI at [Brand]"
• "Our AI presenter, Maya"
• "Animated with AI. Real product, real results."

Give the presenter a name and treat them as a brand character rather than a disclosure liability. Brands that do this consistently — same synthetic presenter, same name, same voice across a year — end up with an asset instead of an asterisk. The audience stops asking after the third video. The AI spokesperson trade-offs breakdown covers where this works and where it backfires.

Placement rules worth writing down: on-screen labels sit in the first three seconds, at readable size, not tucked under the caption fold. A label nobody sees is a label that will be described later as "buried."

Who owns each decision

Policies fail on ownership, not content. Assign three roles by name:

• Policy owner (usually marketing ops or brand): maintains the tier table and the platform appendix, reviews quarterly.
• Tier assigner (the creator): assigns a tier to every asset at creation time, in the same field where they log the model used.
• Escalation owner (brand lead, with counsel on call): rules on anything ambiguous within one business day. The SLA matters more than the seniority — a slow escalation path guarantees people will guess instead.

Add one more rule that prevents most trouble: tier is assigned before generation, not after. If a creator has to declare "this will be a tier 2 asset" while briefing it, they design the label into the storyboard rather than bolting it on after a client has already fallen in love with the cut.

A policy skeleton you can adapt

Keep it under two pages. Longer policies get summarized wrongly in Slack.

1. Purpose — one paragraph on trust, not compliance theater.
2. Scope — which teams, which channels, which asset types. Include sales decks and internal comms; that's where undisclosed synthetic footage quietly leaks out.
3. The tier table — as above, with your own examples.
4. Disclosure standards — exact approved label wordings, placement, and platform toggles.
5. Consent requirements — for cloned voices and digital twins, referencing your consent file process from the legal and licensing basics.
6. Prohibited uses — fake testimonials, real people without consent, fabricated results, competitor likenesses.
7. Record-keeping — model, prompt, tier, approver, publish date.
8. Review cadence — quarterly, with named owner.

Ship version one at 80% confidence. A policy in use gets corrected by reality; a perfect policy in draft protects nobody.

FAQ

Is AI disclosure legally required for marketing content?

It depends on jurisdiction and content type, and it's tightening. Requirements are strongest around realistic depictions of real people, political content, and regulated claims in health and finance. Platform terms often require labeling even where local law doesn't, and those terms are enforced far more frequently than statutes.

Does labeling content as AI-generated reduce reach or engagement?

We haven't seen a consistent penalty from using platform AI toggles honestly, and the downside risk of being labeled by the platform — or called out by a viewer — is far larger. What does hurt is defensive, legalistic wording. Confident framing that names your synthetic presenter performs closer to unlabeled content in the accounts I've watched.

Do we need to disclose AI-generated B-roll or background music?

Generally no. Tier 0 covers AI used as a production tool where nothing is being represented as a real event or real person. Generated cutaways, music beds, upscaling, and auto-captions sit here. Over-labeling dilutes the labels that matter.

Who should own the AI disclosure policy — legal or marketing?

Marketing ops should own the document and the tier table; legal should own the prohibited list and the escalation ruling. Legal ownership of the whole thing produces a policy nobody reads; marketing ownership without a legal veto produces one nobody trusts.

How do we handle AI disclosure for a synthetic brand presenter used long-term?

Name them, introduce them explicitly in early videos, keep a persistent on-screen "AI presenter" credit, and never let them make first-person claims about using the product. Consistency is what converts the disclosure from a warning into brand-consistency: the same face, voice, and label treatment every time.

Write the tier table this week, then encode it where the work happens — build your approved presenter, voice, and label overlay into a reusable Versely workflow so tier 2 assets ship with the label already burned in, and pair it with the trust playbook in brand trust and AI transparency.