AI Video and Platform Guidelines: A Creator Compliance Playbook

If a stranger watched your newest synthetic clip with the sound off — no caption, no context — would they know within two seconds that no camera was involved and no real person appears? That question predicts publishing outcomes better than any slider inside a generation tool. Lighting, camera moves, sound design: that is craft. Whether the craft is ever seen depends on a much duller set of decisions, most of which happen before the first frame renders.

Start with a sentence you can defend out loud

Before you open a tool, write one sentence per video and pin it to the project. The shape is always the same: what is synthetic, who is not real, and what the viewer is actually looking at.

• "Stylized animation, AI-assisted motion, no real people."
• "Synthetic presenter voiced by a licensed model voice; the brand is fictional."
• "Real actor, licensed likeness, synthetic background plate."

If you cannot finish the sentence in one breath, the concept is ambiguous, and ambiguity is what turns a review into a takedown. The sentence is not marketing copy. It is the source of three outputs: the on-screen label, the first line of the caption, and the reply you send if somebody questions the post. One sentence, three uses, zero improvisation at eleven at night.

What platform reviewers actually screen for

Short-form platforms are risk systems, not film critics. Enforcement mixes automated classifiers, keyword signals, viewer reports and human reviewers, and at the volume these systems operate, most decisions take seconds. Nobody is evaluating your dolly shot. They are triaging for harm and deception, and in practice synthetic video gets sorted into four buckets.

Real people without consent. A recognizable face or voice used without permission, or an invented character that reads as a specific living person.

False evidence. Footage that could be mistaken for documentation of a real event, place, quote or product behavior.

Restricted or harmful promotion. Synthetic scenes selling regulated goods, health outcomes, financial promises, or adult themes reachable by minors.

Manipulative framing. Cuts and captions engineered so the viewer believes a synthetic event happened, even if a small label exists somewhere outside the frame of attention.

The consequences are uneven, and that unevenness should drive your concept decisions more than any tool comparison. A heavily stylized cartoon rarely earns a second look. Photoreal synthetic footage of a named public figure almost always does — even when the intent is satire, even when the label is present, even when the likeness is only "close." Build your ideas on the safe side of that asymmetry and your publishing rhythm stops depending on luck.

There is a fifth, softer signal worth naming: contradiction. A caption that says "AI-generated" while the script says "I filmed this myself" is worse than no label at all, because it proves the creator knew and chose to blur it. Reviewers are trained to notice that. Audiences are faster at it than anyone expects, and comment threads now do the enforcement work for free — a clip that feels evasive gets reported, quoted and re-uploaded with the accusation attached.

Disclosure that reads as a fact, not a flourish

Most disclosure failures are one of two mistakes: hiding the label, or contradicting it.

Where the label belongs

• On screen, early. For realistic synthetic footage, plain words in the first two seconds. Not a glyph, not a logo, not a whisper in the corner.
• In the first caption line. Before the follow prompt, before the hashtags, before the emoji wall.
• In the audio. One spoken sentence in the opening five seconds covers people who watch muted-then-unmuted and anyone who reposts a screen recording.
• In the platform's own synthetic-content setting. When a toggle exists, use it. It travels with the post in ways captions do not.
• In your project notes and file metadata. Not for the audience — for you, six months later, when someone asks.

Weak caption, strong caption

Same video, two captions:

• Weak: new drop, link in bio, #ai #fyp #viral
• Strong: Synthetic visuals and an AI voice. No real person shown. Concept film for a fictional brand.

The strong version is boring, specific and nearly impossible to misread. Boring is the goal. A disclosure is a fact label, not a creative flourish, and the moment it starts performing it stops informing.

The opposite failure exists too. Bolting a synthetic-media warning onto every frame of an obvious cartoon trains viewers to ignore warnings, which is exactly the outcome nobody wants. Spend the heavy labeling on content a reasonable person could mistake for reality.

Small tells that invite reports

Labels are not the only authenticity signal. Viewers report what feels off before they can explain why, and these details do the work:

• Faces that drift toward someone recognizable across a series of clips.
• Captions that do not match the synthetic voice word for word.
• Signage, packaging or uniforms appearing in the background of a "generic" scene.
• Watermarks or interface fragments from another tool visible for a frame or two.
• Motion that breaks at the loop seam, or shadows that disagree with the light source.

None of these is a policy violation on its own. Together they push a viewer from "this is a cool animation" to "something is being hidden from me," and that shift is what generates reports.

Tier your concepts before you spend an afternoon rendering

Not every idea deserves the same process. Tiering keeps you fast on safe work and slows you down exactly where it matters.

Tier	What it looks like	Risk	Treatment
0	Abstract textures, landscapes, product macro shots, no people	Very low	Standard edit, no special labeling
1	Stylized characters, animation, obvious fantasy	Low	Light note if it reads photoreal
2	Fictional synthetic presenter, invented persona	Medium	Clear disclosure, no borrowed identity
3	Real public figure, real event, recreated location	High	Written consent or reframe; expect scrutiny
4	Anything that could pass as evidence of a real event	Severe	Do not publish

The tier test is a communication tool as much as a compliance one. Picture a client asking for a "founder announcement" video. A tier 2 solution is an invented presenter with a clear synthetic label. A tier 3 solution is the actual founder, filmed or synthesized with written consent covering AI generation and distribution. A tier 4 solution does not exist — if the deliverable depends on viewers believing a synthetic event happened, the brief is the problem, not the render.

If you cannot explain a concept's tier to a collaborator in one sentence, it is too ambiguous to produce responsibly. Say that early, before anyone has booked a week of editing.

Prompt hygiene: reduce risk at the source

Compliance is mostly a prompt-box discipline. Fix it there and the timeline becomes simple.

Describe behavior, not identities

A middle-aged news anchor in a navy blazer, neutral studio lighting is workable. Looks like a specific famous anchor is not — and naming real people in prompts usually produces unusable output anyway, because models either refuse or drift into caricature. Build characters from traits: age range, wardrobe, posture, energy, speech rhythm. Keep a character sheet so consistency comes from description instead of a borrowed face.

A comparison makes the difference concrete. Instead of handsome actor in his thirties, resembles a known celebrity, write male presenter, late thirties, close-cropped hair, charcoal knit polo, relaxed shoulders, speaks in short sentences, neutral gray studio backdrop, medium shot at eye level. The second version survives a re-render, a style change and a client note without dragging a likeness along with it.

Keep brand marks out of frame

Logos arrive by accident — a soda can on a desk, a sneaker with a swoosh, a storefront sign behind a street scene. Add explicit negatives to every prompt (no logos, no readable signage, no branded packaging) and check each generated clip at full resolution rather than in a thumbnail grid. A single frame of someone else's trademark inside a paid campaign is a conversation nobody wants to have.

Background text deserves its own pass. Synthetic signage often resolves into readable nonsense, which looks sloppy and can accidentally resemble an existing business. Blur it, replace it, or regenerate the plate.

Save the prompt with the output

Every published clip should be traceable to the text that made it. Store prompt, seed, model name and date alongside the file. When a question arrives — from a client, a platform, or your own memory — you answer in seconds instead of reconstructing the week. A prompt library organized by project rather than by mood doubles as that paper trail.

Consent, likeness, voice and music

This is the least glamorous part of AI video and the part that decides whether a campaign can run at all.

• Likeness release. If a synthetic presenter is built from a real person's photos, get a written release covering AI synthesis, derivative frames and distribution channels. "It was just a reference image" is not a defense.
• Voice consent. Cloning a voice needs explicit permission with a defined scope and, ideally, an expiry date.
• Music licensing. A track cleared for organic posting is not automatically cleared for paid amplification. Check both, and keep the license text with the asset.
• Talent and contractor clauses. Update agreements so AI-generated derivatives of a performance are covered before the shoot, not after the first render.
• Minors. Extra consent, extra caution, no exceptions.
• Stock and public-domain sources. Keep the license text in the same folder as the output.

One folder per project — brief, releases, prompts, outputs, final disclosure wording — makes you look organized and, more usefully, keeps decisions consistent when three people touch the same series.

A six-pass workflow for AI video that survives review

Structure beats inspiration when you publish weekly.

Pass 1: brief and claim audit

Write the concept in three lines and assign its tier. Then list every factual or implied claim in the script — health, money, safety, performance — and delete or soften anything you cannot support. One persuasive sentence outperforms five shaky ones, and it is dramatically easier to defend.

Pass 2: stills and storyboard

Lock look, wardrobe and framing before generating motion. Use an AI image generator to produce key frames, then approve the stills as a set. Fixing a face on a still takes seconds; fixing a drifting face across twelve clips takes an afternoon and a re-render.

Pass 3: batch generation and full-resolution review

Generate in small groups, review at full size, keep only what passes. Batch review catches recurring problems early: hands, floating signage, text that resolves into gibberish, a face that has quietly become someone specific. Delete aggressively here. Storage is cheap; a re-shot campaign is not.

Pass 4: assembly, sound and captions

Edit to a rhythm rather than to clip length. Captions must match the synthetic voice word for word — mismatched captions are one of the most common authenticity red flags viewers cite. Keep music licensed and mixed low enough that the spoken disclosure stays audible on a phone speaker in a noisy room.

Pass 5: disclosure and metadata

Apply the exact wording you wrote in Pass 1. Add on-screen text if the tier calls for it, set the platform's synthetic-content flag, and put the plain-language note in both the caption and the project log. If a brand has approval rights, this is the version they sign off on.

Pass 6: publish and log

Record the URL, date, tier and disclosure used. That log is your fastest defense if a post is questioned and your best guide to what to repeat. Teams publishing at volume standardize openings, caption blocks and label placement with video templates — consistency is a compliance feature, not just a branding one.

Sponsored content adds a second rulebook

Advertising layers endorsement rules on top of platform policy. Paid partnership labels have to be visible, not buried in a hashtag pile. A claim delivered by a synthetic voice is still a claim, and "the model said it" has never worked as a defense. Habits that hold up:

• Route health, finance and safety claims through the brand's reviewer before generation, not after.
• Keep the ad disclosure in the first frame and in the caption's opening line.
• Avoid synthetic depictions of real endorsers unless you hold explicit, documented rights.
• Get written approval of the final cut and archive it with the project folder.
• Agree who owns the output if the partnership ends; unclear ownership resurfaces later.

Regulators in most markets publish plain-language guidance on endorsement clarity and on synthetic media in advertising. Read the version relevant to your audience once. It is a twenty-minute investment that pays out across every campaign that follows.

When a post gets flagged: a calm response plan

1. Do not delete reflexively. Read the notice and identify which bucket the platform is pointing at.
2. Compare the notice to your disclosure. Was the label visible in the first two seconds? Did the caption contradict it?
3. Fix the ambiguity, not just the label. Re-cut, re-render or reframe; a bigger warning on a misleading edit does not solve the underlying issue.
4. Appeal with documentation if the enforcement looks mistaken: release forms, prompt log, a screenshot of the disclosure as published.
5. Log the cause and update your checklist so the same mistake does not repeat across an entire series.

Treat a flag as data. Most repeat offenders are not reckless; they are solving the same problem from scratch every week because nobody wrote the fix down.

FAQ

Does every AI-generated video need a label?

No. Obvious animation, abstract visuals and product macro shots rarely need one. Label whenever a reasonable viewer could mistake the footage for real people, real events or real documentation — and when you do, label early.

Can I use a real person's face if I have their permission?

Usually yes, with written consent that explicitly covers AI synthesis, derivative frames and the channels you plan to publish on. A verbal yes is not consent, and "they posted it publicly" is a story that ends badly.

What about satire and parody?

Satire is often protected expression, but platform policy and disclosure duties still apply. Keep the label, avoid framing that resembles evidence, and be careful placing real people in contexts a viewer could mistake for fact.

Is an AI voiceover allowed in advertising?

Yes, with licensed or consented voice use, visible ad disclosure and claims the brand can substantiate. Cloning a recognizable person's voice requires explicit permission, ideally with scope and expiry written down.

What should I do if a client asks me to hide the AI label?

Decline and offer alternatives: a more stylized treatment, a fictional presenter, or a rewritten claim. Removing the label moves risk onto the client's brand and your reputation, and it rarely buys the performance the client expects.

How do I keep a series consistent without reusing a real face?

Build a character sheet — age range, wardrobe, posture, vocal cadence — and describe it in every prompt. Consistency from description keeps you in the fictional tier and out of likeness disputes.

Do I need to keep old prompts forever?

Keep them as long as the video stays published plus a reasonable buffer. They cost almost nothing to store and they turn a defensive conversation into a ten-second lookup.

Build the next video on a policy you can defend

Compliant AI video is not slower to make. It is planned in a different order: concept, tier, consent, prompt hygiene, disclosure, publish. Once that sequence is a habit, the creative work gets easier, because you stop guessing which ideas are safe to invest in.

Start small. Take one concept, run it through the tier table, write your one-sentence disclosure, then render it with the Orelon AI video generator. When the workflow feels natural, browse the Orelon blog for more production patterns. Cinematic ideas in motion are worth publishing — and worth publishing in a way that stays up.