Content filtering for model inference in Azure AI services

Azure AI model inference in Azure AI Services includes a content filtering system that works alongside core models and it's powered by Azure AI Content Safety. This system works by running both the prompt and completion through an ensemble of classification models designed to detect and prevent the output of harmful content. The content filtering system detects and takes action on specific categories of potentially harmful content in both input prompts and output completions. Variations in API configurations and application design might affect completions and thus filtering behavior.

The text content filtering models for the hate, sexual, violence, and self-harm categories were trained and tested on the following languages: English, German, Japanese, Spanish, French, Italian, Portuguese, and Chinese. However, the service can work in many other languages, but the quality might vary. In all cases, you should do your own testing to ensure that it works for your application.

In addition to the content filtering system, Azure OpenAI Service performs monitoring to detect content and/or behaviors that suggest use of the service in a manner that might violate applicable product terms. For more information about understanding and mitigating risks associated with your application, see the Transparency Note for Azure OpenAI. For more information about how data is processed for content filtering and abuse monitoring, see Data, privacy, and security for Azure OpenAI Service.

The following sections provide information about the content filtering categories, the filtering severity levels and their configurability, and API scenarios to be considered in application design and implementation.

Content filter types

The content filtering system integrated in the Azure AI Models service in Azure AI Services contains:

• Neural multi-class classification models aimed at detecting and filtering harmful content. These models cover four categories (hate, sexual, violence, and self-harm) across four severity levels (safe, low, medium, and high). Content detected at the 'safe' severity level is labeled in annotations but isn't subject to filtering and isn't configurable.
• Other optional classification models aimed at detecting jailbreak risk and known content for text and code. These models are binary classifiers that flag whether user or model behavior qualifies as a jailbreak attack or match to known text or source code. The use of these models is optional, but use of protected material code model might be required for Customer Copyright Commitment coverage.

Risk categories

^* If you're an owner of text material and want to submit text content for protection, file a request.

Text content

Image content

Prompt shield content

Configurability

Models deployed to Azure AI Services include default safety settings applied to all models, excluding Azure OpenAI Whisper. These configurations provide you with a responsible experience by default.

Certain models allow customers to configure content filters and create custom safety policies that are tailored to their use case requirements. The configurability feature allows customers to adjust the settings, separately for prompts and completions, to filter content for each content category at different severity levels as described in the table below. Content detected at the 'safe' severity level is labeled in annotations but is not subject to filtering and isn't configurable.

¹ For Azure OpenAI models, only customers who have been approved for modified content filtering have full content filtering control and can turn off content filters. Apply for modified content filters via this form: Azure OpenAI Limited Access Review: Modified Content Filters. For Azure Government customers, apply for modified content filters via this form: Azure Government - Request Modified Content Filtering for Azure OpenAI Service.

Content filtering configurations are created within a resource in Azure AI Foundry portal, and can be associated with Deployments. Learn how to configure a content filter

Scenario details

When the content filtering system detects harmful content, you receive either an error on the API call if the prompt was deemed inappropriate, or the finish_reason on the response will be content_filter to signify that some of the completion was filtered. When building your application or system, you want to account for these scenarios where the content returned by the Completions API is filtered, which might result in content that is incomplete. How you act on this information is application specific. The behavior can be summarized in the following points:

• Prompts that are classified at a filtered category and severity level will return an HTTP 400 error.
• Nonstreaming completions calls won't return any content when the content is filtered. The finish_reason value is set to content_filter. In rare cases with longer responses, a partial result can be returned. In these cases, the finish_reason is updated.
• For streaming completions calls, segments are returned back to the user as they're completed. The service continues streaming until either reaching a stop token, length, or when content that is classified at a filtered category and severity level is detected.

Scenario: You send a nonstreaming completions call asking for multiple outputs; no content is classified at a filtered category and severity level

The table below outlines the various ways content filtering can appear:

Example request payload:

{
    "prompt":"Text example", 
    "n": 3,
    "stream": false
}

Example response JSON:

{
    "id": "example-id",
    "object": "text_completion",
    "created": 1653666286,
    "model": "davinci",
    "choices": [
        {
            "text": "Response generated text",
            "index": 0,
            "finish_reason": "stop",
            "logprobs": null
        }
    ]
}

Scenario: Your API call asks for multiple responses (N>1) and at least one of the responses is filtered

Example request payload:

{
    "prompt":"Text example",
    "n": 3,
    "stream": false
}

Example response JSON:

{
    "id": "example",
    "object": "text_completion",
    "created": 1653666831,
    "model": "ada",
    "choices": [
        {
            "text": "returned text 1",
            "index": 0,
            "finish_reason": "length",
            "logprobs": null
        },
        {
            "text": "returned text 2",
            "index": 1,
            "finish_reason": "content_filter",
            "logprobs": null
        }
    ]
}

Scenario: An inappropriate input prompt is sent to the completions API (either for streaming or nonstreaming)

Example request payload:

{
    "prompt":"Content that triggered the filtering model"
}

Example response JSON:

"error": {
    "message": "The response was filtered",
    "type": null,
    "param": "prompt",
    "code": "content_filter",
    "status": 400
}

Scenario: You make a streaming completions call; no output content is classified at a filtered category and severity level

Example request payload:

{
    "prompt":"Text example",
    "n": 3,
    "stream": true
}

Example response JSON:

{
    "id": "cmpl-example",
    "object": "text_completion",
    "created": 1653670914,
    "model": "ada",
    "choices": [
        {
            "text": "last part of generation",
            "index": 2,
            "finish_reason": "stop",
            "logprobs": null
        }
    ]
}

Scenario: You make a streaming completions call asking for multiple completions and at least a portion of the output content is filtered

Example request payload:

{
    "prompt":"Text example",
    "n": 3,
    "stream": true
}

Example response JSON:

 {
    "id": "cmpl-example",
    "object": "text_completion",
    "created": 1653670515,
    "model": "ada",
    "choices": [
        {
            "text": "Last part of generated text streamed back",
            "index": 2,
            "finish_reason": "content_filter",
            "logprobs": null
        }
    ]
}

Scenario: Content filtering system doesn't run on the completion

Example request payload:

{
    "prompt":"Text example",
    "n": 1,
    "stream": false
}

Example response JSON:

{
    "id": "cmpl-example",
    "object": "text_completion",
    "created": 1652294703,
    "model": "ada",
    "choices": [
        {
            "text": "generated text",
            "index": 0,
            "finish_reason": "length",
            "logprobs": null,
            "content_filter_result": {
                "error": {
                    "code": "content_filter_error",
                    "message": "The contents are not filtered"
                }
            }
        }
    ]
}

Next steps

• Learn about Azure AI Content Safety.
• Learn more about understanding and mitigating risks associated with your application: Overview of Responsible AI practices for Azure OpenAI models.
• Learn more about how data is processed with content filtering and abuse monitoring: Data, privacy, and security for Azure OpenAI Service.