Where the content filter actually sits
A content filter is a separate classifier the host runs on your prompt, the model’s reply, or both, and when it fires you get a block instead of an answer. It is not the model refusing, and turning it off does not stop an aligned model from declining. Telling the two apart is the first step to fixing either.
Filter versus refusal
| Content filter | Model refusal | |
|---|---|---|
| What does it | A classifier run by the host around the model | The model itself, from its alignment training |
| What you get back | A blocked response, an error or an empty reply with a safety reason | A normal reply that declines in words |
| How to change it | Host settings, where the host allows it | A different model: uncensored or abliterated |
Which APIs filter by default
- Azure OpenAI runs content filters on both prompt and completion by default, across hate, sexual, violence and self-harm. Relaxing them requires an approved application as a managed customer. See Azure OpenAI alternatives.
- The Gemini API and Vertex AI apply safety settings per harm category with adjustable thresholds. A blocked reply comes back with a safety finish reason rather than text. See Gemini and Vertex.
- AWS Bedrock offers Guardrails as an optional layer you attach, on top of each vendor's own model policy. See Bedrock.
- GitHub Models applies Azure AI Content Safety on top of whichever model you pick. See GitHub Models.
- The OpenAI API offers moderation as a separate endpoint you can call yourself, but its models are trained to refuse, so the refusal is in the model rather than in a filter.
What an unfiltered API actually needs
Two things, and most "no filter" claims only deliver the first. No classifier in front of the model, so nothing is blocked on the way in or out. And a model that does not refuse on its own, which means an abliterated or uncensored build rather than a frontier model with a filter switched off.
Wild West API runs no moderation layer: it prices the call, holds the money against the key's cap and passes the request to the model. The models it sells are uncensored, and one is abliterated. What comes back is what that model does.
Recognising a filter from the response
A filter usually leaves a trace a refusal does not: an HTTP error naming a policy, a finish_reason of content_filter or a safety category, or an empty message. A refusal arrives as an ordinary reply that politely says no. If the reply is words, changing the filter settings will not help; changing the model will.
FAQ
Is there an LLM API with no content filter?
Yes. Self-hosted models have none, and several hosts serve uncensored models without a moderation layer. Check that the models are uncensored too, or the model will still refuse.
Can I turn off the Azure OpenAI content filter?
Only partly, and only after Microsoft approves an application for modified content filtering as a managed customer. Most accounts cannot turn it down.
Why does my API return an empty response?
Often a safety filter: the Gemini API, for example, returns a safety finish reason instead of text when a category trips. Check the finish reason before assuming the model failed.