There are usually two of them and they behave differently. An input filter reads the prompt and refuses before any compute is spent, which is why a rejection can be instant. An output filter inspects the finished frames and suppresses them, which is why the other kind of rejection arrives after the wait, sometimes as a blank or black result rather than an error.
Filters are classifiers, so they are wrong in both directions. False positives are the common annoyance — anatomical terms in a medical brief, violence in a sports context, a brand name that collides with something else. False negatives exist too, which is why a provider's terms, not the filter, define what you are allowed to publish.
Where a tolerance setting is exposed, it moves a threshold within limits the provider set; it is not an off switch, and the provider's policy still governs regardless of the value.
In practice
- Instant refusals are usually prompt-level; late blanks are usually output-level.
- Rephrasing around a trigger word often clears a false positive without changing the brief.
- Filters differ per provider, so the same prompt can pass on one model and fail on another.
The mistake to avoid
Reading a blocked generation as a bug and retrying identically. Retries of a prompt-level refusal fail the same way; the wording has to change.
Related terms
Negative prompt
A negative prompt lists what you do not want in the output, steering the generation away from those terms instead of towards them.
Prompt
A prompt is the written instruction a generative model reads to decide what to make — the one input almost every model requires.
Watermark
A watermark is a mark applied to generated media to identify where it came from — either visible on the picture or embedded invisibly in the file.
Prompt expansion
Prompt expansion is a provider-side step that rewrites your short prompt into a longer, more detailed one before the generator ever sees it.
Seed
A seed is the number that decides the random starting noise for a generation, so the same seed with the same settings reproduces the same output.
The all-in-one AI studio for creators. 60+ models for video, image, voice, music and lipsync in a single app.