X's manipulated media policy and AI video
X's synthetic media rule turns on deception plus potential harm, not on whether a model made the frames. Here is the two-part test and what actually trips it.
The most useful thing to understand about X's synthetic and manipulated media policy is what it does not ask. It does not ask whether a model generated the frames. It does not ask whether you used a diffusion pipeline or a camera. It asks whether the media could deceive, and whether that deception could cause harm — and where both are true, the post may be labelled or removed.
That's a two-part test, and both parts have to land. Which means a great deal of AI video is nowhere near the policy, and a small amount of entirely un-generated video is squarely inside it. Creators who file this under "AI rules" get the risk map backwards.
The test has two conditions, not one
Write it out as a grid, because the four cells behave completely differently:
| No potential harm | Potential harm | |
|---|---|---|
| Not deceptive | Fine. Most creative work. | Fine on this rule. Other rules may apply. |
| Deceptive | Usually fine on this rule — a stylised fake nobody could act on. | This is the cell the policy is for. |
The bottom-right cell is the entire policy. Everything else is out of scope for this particular rule, which is not the same as saying it's out of scope for every rule — impersonation, platform manipulation, and the monetization standards all sit elsewhere and have their own tests.
Neither condition is a property of the video file. Deception depends on presentation: the same clip is deceptive with one caption and clearly fictional with another. Harm depends on subject: a fabricated product launch and a fabricated emergency are the same technical artifact with different consequences.
The artifact you are assessing is the video plus the caption plus the account plus what a reasonable viewer would conclude. That's inconvenient for a checklist and it's the honest reading of a rule written around effect rather than technique.
The AI video categories that actually trip it
X's own policy page is the authority on what X enforces, and it's worth reading directly. For a working taxonomy of what "could deceive" looks like in generated video, YouTube's published disclosure guidance is a useful scaffold, because it enumerates the three cases plainly — making a real person appear to say or do something they didn't, altering footage of a real event or place, and generating a realistic scene that never occurred.
Those three are YouTube's categories, not X's. But they map cleanly onto the deception half of X's test, and every one of them can pair with harm:
A real person saying or doing something they didn't. The highest-risk category, because deception and harm arrive together by default. A synthetic clip of a named public figure making a statement is deceptive on its face and harmful in almost any context where someone might act on it. A disclaimer in the fourth line of the caption does not solve this.
Altered footage of a real event or place. Real source footage, changed. A crowd made larger, a building damaged, weather added, a location swapped. The deception is subtle precisely because the base material is genuine, and the harm scales with how newsworthy the underlying event is.
A photorealistic scene that never occurred, presented as documentation. Fully generated, no real source, but framed as something that happened. The generation itself is neutral. The framing is what puts it in the bottom-right cell.
Two more that come up constantly on AI-first accounts and belong on the list:
A synthetic spokesperson presented as a real employee, customer, or expert. Testimonial and UGC formats live here. A generated presenter reading a script is ordinary production. A generated presenter introduced as "one of our customers" is a fabricated endorsement, and the harm is commercial rather than civic but it's still harm. If you're building in this space, the wider AI spokesperson trade-offs post covers the disclosure side.
Armed conflict, and other crisis footage framed as documentation. Generated footage of a disaster, an outbreak, a safety incident, or a war. The deception-plus-harm test already covers this when the clip is presented as a record. Separately, X's content monetization standards require a clear disclosure on any AI-generated video depicting armed conflict, and they require realistic video-game depictions of armed conflict to be labelled synthetic. In March 2026, X's head of product announced a 90-day Creator Revenue Sharing suspension for unlabeled AI-generated armed-conflict video, with a permanent ban on a further violation — reported by TechCrunch and The Guardian. Treat the disclosure duty as current. Do not treat that 90-day clock as a general unlabeled-AI penalty, or as evidence of a composer toggle.
What clearly doesn't trip it
The list on the other side is longer, and it covers most of what a working creator actually produces.
Animation, stylised worlds, and anything that reads as illustration rather than photography. Beauty and colour work. Special effects on your own footage. A cloned version of your own voice. Captions and dubbing. Product renders of your own product. Explainer footage that is obviously constructed. Fictional characters in fictional scenes. None of these deceive anyone about a real event or a real person, which means the first condition never fires and the second never gets evaluated.
YouTube's disclosure guidance makes the same carve-outs explicit — animation and fantasy, beauty filters, colour work, special effects, cloning your own voice, captions, audio repair, and production assistance. The reasoning transfers even where the exact rule doesn't: content that doesn't purport to document reality isn't in the deception business.
The ordinary catalogue on a generated-content account — b-roll, explainers, product video, stylised shorts, generated characters in generated worlds — is not where this policy risk lives. The risk concentrates in a narrow band: real people, real events, real places, and anything framed as documentation.
An operating rule that scales
Most teams do better with one production rule applied upstream:
Never present generated footage as a record of something that happened.
Generate whatever you want. Style it however you want. Don't frame it as documentation of a real person, event, or place unless it is one. Every category above violates this rule; nothing in the exempt list does.
Three implementation notes:
The caption is part of the asset. A clip that's obviously fictional in isolation becomes deceptive with "footage from this morning" above it. Whoever writes captions needs the same rule the person generating has. If posting or scheduling to social is handled by someone other than the person who made the video, this is where the rule leaks.
Real people need a separate gate. Likeness is not just a policy question. It's a consent and rights question that survives whatever the platform decides, and the likeness release side of it sits outside this policy entirely. A post that wasn't removed is not a post that was cleared.
Disclosure helps, but it isn't a licence. A visible "generated with AI" line reduces the chance a viewer is deceived. It doesn't neutralise harm. Writing a disclosure line nobody scrolls past is worth doing; it is not a way to publish a fabricated statement from a real person. X's help pages do not describe a post-level "made with AI" control, so disclosure on X is something you write. For AI-generated armed-conflict video, that written line is a published monetization requirement.
How this compares across platforms
If you publish the same asset to several destinations, the tests are not identical and one clearance does not cover all of them.
| Platform | What the rule keys on |
|---|---|
| X | Deception plus potential harm; label or removal |
| YouTube | Disclosure required for realistic altered/synthetic content in three named cases; disclosure itself doesn't affect reach or monetization |
| Meta surfaces | "AI Info" labelling from metadata or self-disclosure; the sharper constraint is originality, not provenance |
Labelling is cheap and near-universal, and it is not the thing that costs you distribution. Across the platforms that publish rules, the expensive failures are originality and mass production, not disclosure. The cross-platform labelling checklist is the version you can work from, including the automated labelling paths that fire on Meta surfaces whether or not you self-declare.
For the production layer, this mostly means keeping provenance clean and your own: generate your own assets rather than sourcing footage of uncertain origin (AI video generator), keep a re-renderable timeline so a caption fix doesn't mean regenerating everything, and understand synthetic media disclosure as a category before you need it.
FAQ
Does labelling my post as AI-generated make it compliant?
Not by itself. Disclosure addresses the deception half of the test — a viewer who knows the footage is synthetic is less likely to be misled. It does nothing about the harm half. A labelled fabrication of a real person saying something they didn't is still a fabrication of a real person.
Is stylised or animated AI video at risk under this policy?
Generally not on this rule. If nobody could mistake it for a record of something that happened, the deception condition doesn't fire. Other rules — impersonation, monetization standards, advertising requirements — apply independently.
What about generated footage of a real place, like a city street?
It depends on framing. A generated street scene as background or establishing shot is ordinary production. The same scene captioned as showing conditions in that city on a given day is altering the record of a real place, which is squarely in the deception column.
Does X have a "made with AI" toggle I should be using?
Nothing on X's published help pages describes a post-level AI toggle. Until X publishes one, disclosure on X is something you write into the post or your profile yourself. For AI-generated videos of armed conflict, the monetization standards already require a clear disclosure whether a toggle exists or not.
X's manipulated media policy page and its content monetization standards are the authoritative text; both change, and the version on X's own help domain is the one that governs. Read it directly before you build a workflow around any summary, including this one.