What is Content moderation? Definition, key points and related tools
Compliance screening of prompts and outputs. Chinese models reject gore, corpses and nudity terms (e.g. Wan's DataInspectionFailed). Whether a blocked request is charged varies: Google does not charge for blocked Veo videos, while Runway's API charges moderated generations the same as successful ones. Keep violence at cause and consequence in storyboards.
Two checkpoints: before and after rendering
Moderation runs at two points. Input screening reads the prompt and any reference image or audio before the model starts. Output screening checks the rendered frames and sound. Vendors expose both: Alibaba Cloud Model Studio returns DataInspectionFailed when input or output data may contain inappropriate content, and IPInfringementSuspect when input data is suspected of IP infringement (Alibaba Cloud); MiniMax returns error 1026 for sensitive input and 1027 for sensitive output (MiniMax); Runway's API moderates both task inputs and outputs and reports SAFETY.INPUT or SAFETY.OUTPUT failure codes (Runway); Google notes that Veo 3.1 sometimes blocks a video because of safety filters or audio processing issues (Gemini API).
The distinction decides the fix: an input block is about words and reference files, an output block about what the staging produced. AI short drama is exposed to these filters because its staple genres (revenge, crime, romance, family conflict) run on violence, intimacy and sometimes children on screen, and one rejected segment holds up the episode it belongs to.
What trips filters in drama work
- Wounds and gore spelled out. Keep violence at cause and consequence; the common-failures explainer has rewrites.
- Real people. OpenAI's Sora API does not generate real people, public figures included, and rejects input images that contain human faces (OpenAI). Runway's API lets developers set how strictly recognizable public figures are handled through a contentModeration setting.
- Minors. Sora's API is limited to content suitable for audiences under 18. Veo 3.1 accepts allow_adult as the sole personGeneration value for image-to-video and reference-image requests, and for every request from EU, UK, Swiss and MENA locations, so child characters in a reference-driven workflow need checking before a season is planned on it.
- Copyrighted characters and music. Sora's API rejects them; Alibaba screens for suspected IP infringement.
- Prompt rewriting. Wan 3.0 turns prompt_extend on by default (Alibaba Cloud), so the text the model receives is not exactly what you wrote. Turning it off makes a block traceable to your own words (see prompt rewriting).
Who pays for a blocked generation
| Vendor | Stated policy | Source |
|---|---|---|
| Google Veo 3.1 (Gemini API) | Not charged if the video is blocked from generating | Gemini API |
| Runway (developer API) | Moderated generations cost the same credits as successful ones; credits are not refunded for SAFETY.INPUT failures; an account with too many moderated requests is suspended | moderation, task failures |
| Kling | Credits refunded when a generation fails; moderation not addressed in the credit policy | Kling |
| SceneMixer | Tasks that fail, are blocked by content moderation or time out upstream are refunded automatically | SceneMixer terms |
| Alibaba Cloud Model Studio, MiniMax | Error codes documented; billing for blocked requests not stated on those pages | Alibaba Cloud, MiniMax |
A pipeline tool inherits the filter of whichever upstream model a tier runs on, so switching tiers changes the filter as well as the price. How refund rules feed into budgets is covered under credit-based billing.
A pre-flight routine
- Lint storyboards before submission for wound and gore terms, nudity, real names and brand characters, and fix them in the storyboard rather than retry by retry.
- Keep reference art synthetic: generated character sheets, no photos of real people.
- Read the error before retrying. After an input block, change wording or the reference file; after an output block, change the staging (framing, distance, what is visible), not synonyms.
- Do not loop the same prompt. Repeated blocks can cost credits on some platforms and, on Runway's API, the account.
- Log blocks per model. If one engine rejects a genre's staple scene, route those segments to another engine.
Rights, likeness and labeling questions sit beside moderation; the copyright and labeling checklist covers them.
Key figures and sources
| Figure | Value | Source | Date |
|---|---|---|---|
| Alibaba Cloud Model Studio block codes | DataInspectionFailed (input or output may contain inappropriate content); IPInfringementSuspect (suspected IP infringement) | www.alibabacloud.com | 2026-09-12 |
| MiniMax moderation error codes | 1026 sensitive input; 1027 sensitive output | platform.minimax.io | 2026-09-12 |
| Veo 3.1 blocked videos | Not charged when blocked; personGeneration limited to allow_adult for image and reference inputs and in EU, UK, CH and MENA | ai.google.dev | 2026-09-12 |
| Runway API moderation | Moderated generations cost the same credits as successful ones; too many moderated requests lead to suspension | docs.dev.runwayml.com | 2026-09-12 |
| Sora API restrictions | No real people or public figures; input images with human faces rejected; copyrighted characters and music rejected | developers.openai.com | 2026-09-12 |
FAQ
Why was a prompt blocked when nothing in it is explicit?
Filters read more than the prompt text: a reference image with a face (rejected by Sora's API), a child character sheet in an image-driven Veo request, a character name that matches a public figure, or wording added by automatic prompt rewriting can all trigger a block. Check whether the error is an input or an output block first.
Do I lose credits when a generation is blocked?
It depends on the vendor. Google does not charge for blocked Veo videos; SceneMixer's terms refund moderation-blocked tasks; Runway's developer API charges moderated generations the same as successful ones; Kling refunds failed generations without mentioning moderation; Alibaba Cloud and MiniMax do not state it on their error-code pages.
Is moderation stricter on some models than others?
It differs by category rather than on one scale: Sora's API excludes real people and faces in input images, Veo restricts people generation to adults in image-driven modes and some regions, and Alibaba screens for IP. A scene blocked on one engine may pass on another, so test a genre's key scenes on the engine before committing a season to it.
Sources
- https://www.alibabacloud.com/help/en/model-studio/error-code
- https://www.alibabacloud.com/help/en/model-studio/wan3-video-generation-api-reference
- https://platform.minimax.io/docs/api-reference/errorcode
- https://docs.dev.runwayml.com/api-details/moderation/
- https://docs.dev.runwayml.com/errors/task-failures/
- https://ai.google.dev/gemini-api/docs/veo
- https://developers.openai.com/api/docs/guides/video-generation
- https://kling.ai/docs/point-policy
- https://scenemixer.com/terms/