Why did I get an unsupported prompt or policy error?

Each AI model provider runs its own content-safety filter on every prompt and source asset. The most common triggers are: real public figures named or visually referenced, sexual content of any kind on most flagship models, graphic violence and gore, hate speech or harassment targets, branded characters or copyrighted IP without context, and weapons or self-harm references. The filter operates on both the prompt text and any uploaded reference images, so a borderline reference image can cause a rejection even with a clean prompt.

The error message names the category broadly (Content policy violation, Safety filter rejection, Unsupported prompt) but does not return the exact rule that fired. Model providers intentionally do not publish a banned-words list — the filter is dynamic and naming words would let users route around it. Treat the rejection as a category signal, not a precise diagnosis.

Policy rejections that happen at the provider after the request was submitted refund credits automatically. The refund posts within seconds and the failed run is excluded from your billable usage. You do not need to file a support request. The Settings, Billing page shows the refund entry alongside the failed run, so you can confirm the credit returned.

Policy rejections that happen before the request is sent — when Martini's own pre-send check catches a clearly disallowed prompt — never deduct credits in the first place. There is nothing to refund. Either way, the financial outcome is the same: you are not charged for a rejected generation. See the failed-generation refunds article for the full credit behavior on errors.

Replace named real people with generic descriptions — a politician becomes a public speaker, a celebrity becomes a person in a similar style. Remove brand names, copyrighted character names, and specific IP references. Reword scene descriptions that could be read as violent or sexual even when innocent — soften combat verbs, remove blood references, drop intimate physical descriptors. Swap the source reference image if it was the trigger.

After rewording, run again. Most legitimate creative briefs can be expressed in a way that passes the filter — the goal is to convey the same scene without naming the specific identifying triggers. If the request is genuinely about a real public figure or branded property, you may need an external license rather than a prompt change.

Different providers calibrate their filters differently. Some video and image models are stricter on public figures and brands; others are more permissive but stricter on violence or sexual themes. If a prompt is rejected on one model, try the same prompt on a sibling model — the rejection often does not transfer. The model picker on every node lets you switch providers without rebuilding the workflow.

Models also differ on whether the safety pass runs once on input or twice (input + output). A two-pass model can return a rejection minutes after submission if the generated frame trips the second pass. This is rare but explains why some long-running video jobs fail late with a policy error rather than failing immediately.