Version: draft-2026-08-25 · Status: ⚠️ DRAFT — not reviewed by a lawyer
This is what moderation actually enforces, written in the same categories the code uses. It is part of the Terms of Service.
These end the account immediately and, where the law requires it, are reported. There is no appeal and no warning.
any form, including generated or stylised.
designed to cause mass casualties, or for attacks on people or infrastructure.
Generation is refused. Repeated attempts escalate — see the ladder below.
Impersonation and deception
public body when it is not.
This is the category that most often catches somebody by surprise. Making an ad "with a celebrity in it" is not a grey area here.
Sexual and adult content
Explicit sexual content, and nudity presented sexually. This is an advertising tool; the platforms it publishes to prohibit the same thing.
Hate and harassment
Content that attacks or demeans people on the basis of race, ethnicity, national origin, religion, disability, sex, gender identity or sexual orientation. Content that targets a specific private individual.
Violence and self-harm
Graphic violence, gore, and anything that encourages, instructs in or glorifies self-harm, suicide or eating disorders.
Illegal goods and regulated claims
The last two are refused because they are the ones that turn a customer's ad into a regulatory problem for the customer.
Intellectual property
Third-party trade marks, characters or brand assets used in a way that implies endorsement or origin.
Not banned, but treated as high-risk and always reviewed by a person before it can be published through the service. Synthetic media in an electoral context is regulated differently in almost every jurisdiction, and the review exists so that a customer does not discover that after publishing.
Before generation. The prompt is checked by rules and by a language model. The two together, because rules alone miss anything phrased carefully and a model alone is inconsistent about the same input.
After generation. Output is sampled — always where the request involved a real face, and at a risk-weighted rate otherwise. Full inspection of every frame of every video is not affordable and would not catch more.
The ladder. Not one strike:
| Response | |
|---|---|
| First refusal | The generation is blocked. Nothing else happens |
| Repeated refusals | Rate limited, and flagged for review |
| A pattern | Generation suspended pending human review |
| Serious violation | Account terminated |
Automated moderation is wrong sometimes, in both directions. The ladder exists so that being wrong once costs a customer one generation rather than their account.
Every rejection has an appeal button. It goes to a person, not back to the model.
We aim to answer within 24 hours; the platform panel tracks the age of the oldest open review and alerts on it at 12 hours, so the target is measured rather than aspirational.
If an appeal succeeds, the block is lifted and the strike is removed from the account.
Your uploaded brand assets, your own product photos, and text you write that is never sent for generation. We check what goes to a provider and what comes back.
We do not read your prompts for any purpose other than this check, and the moderation record stores the decision and the category — not a copy of your content.
{{ABUSE_EMAIL}}. Include the ad's URL or the sign-off link if you have it. Reports about content involving minors are handled first, ahead of everything else.
document wise before it is contractually required?
fast the rules are changing?
jurisdiction we sell into?