Azure AI Content Safety Frequently Asked Questions

We believe it is our responsibility to keep our content safety models up to date with the latest attack patterns. We continuously update our models and monitor behavior and production traffic to ensure each model version is working as expected. If our content safety models are not performing well for your workload, you can provide representative data so we can improve our models specifically for your use case.

We continuously improve our content safety models to improve quality and accuracy. If your prompt or completion is incorrectly categorized as harmful or unsafe content, or if harmful or unsafe content is not being blocked as expected, you can submit data to us for the purpose of improving our content safety models.

To submit incorrectly categorized data, please reach out to Azure AI Content Safety Support at contentsafetysupport@microsoft.com. Alternatively, you can open a support request to get immediate assistance.

Create a support request to submit a ticket:

• For Issue type, select Technical.
• For Subscription, select your subscription.
• For Service, select All services, then select Content Safety.
• For Summary, add a description of your issue.

Microsoft has the lowest latency filters of any major LLM provider.

If latency is still an issue, we recommend Asynchronous filtering. In this case, content filters are run asynchronously, and completion content is returned immediately with a smooth token-by-token streaming experience. No content is buffered, which allows for a fast streaming experience with zero latency associated with content filtering. However, unsafe content might be briefly exposed before filtering is complete.

The model requires time to iterate and retrain. If you create a support ticket through the Azure portal, a customer support engineer assists you in tracking the status and updating you on progress. However, the product team encourages you to mitigate issues more quickly on your own using our recommended tools, such as the blocklist. Refer to Mitigate false results in Azure AI Content Safety.

For instructions on how to start submitting text and images to Azure AI Content Safety and view model responses and results, visit the Microsoft Foundry portal page. Use the sidebar to navigate to the Guardrails + controls page.

Our content harm classification models support the moderation of text, images, and multimodal content (images with text + OCR). The protected material, prompt shields, and groundedness detection models work with text content only.

Azure AI Content Safety generally charges by volume. For example, the Image API is priced based on the number of images submitted. The Text API is billed for the number of text records submitted to the service. However, each model has its own corresponding rate. See the Azure pricing page for more information about pricing tiers.

No. The Azure AI Content Safety API returns classification metadata based on model outputs. Our results tell users whether material across various classes (sexual, violence, hate, self-harm) is present in input content, through either a returned severity level (such as in the Text API) or binary result (such as in Prompt Shields API).

As the user, you use those results to inform appropriate enforcement actions—such as automatically tagging or removing certain content—based on your own policies and practices.

Service usage is throttled if you reach the transaction limit on the Free tier.

Refer to the overview for the current RPS limits for each API. To request an increase in RPS, email us with justification and an estimated traffic forecast.

To decide whether to filter at severity level 0, 2, 4 or 6, we strongly encourage you to consult the descriptive examples of each severity level listed in Harm Categories. Generally, text that scores at severity level 0 might reference a sensitive topic but may be benign. Deciding whether to moderate at level 2 or 4 depends on your community guidelines and sensitivity around controversial topics.

Overall, it depends on your datasets and your own policies and risk tolerance. We recommend starting at a severity level of 4 for new projects. You can then adjust your severity level based on model performance.

No. Currently, text moderation tasks are limited to 10k-character submissions. You can, however, split longer text content into segments (for example, based on punctuation or spacing) and submit each segment as related tasks to the Content Safety API.

See the Azure Content Moderator to Azure AI Content Safety migration guide.

Azure Content Moderator is deprecated as of March 2024. Microsoft recommends that Azure Content Moderator users migrate to Azure AI Content Safety because:

• Azure AI Content Safety offers more accurate and granular detection of harmful content in text and images using state-of-the-art AI models.
• Azure AI Content Safety supports multilingual content moderation (see Language support).
• Azure AI Content Safety enables responsible AI practices by monitoring both user-generated and AI-generated content.

The main differences between the two services are:

• Azure Content Moderator uses a binary classification for each content flag (such as profanity or adult), while Azure AI Content Safety uses multiple severity levels for the different classes (sexual, violent, hate, and self-harm).
• Azure AI Content Safety supports multilingual content moderation (see Language support), while Azure Content Moderator's AI classifiers only support English.
• Azure Content Moderator has a built-in term list and a custom term list feature, while Azure AI Content Safety doesn't have a built-in term list but relies on advanced language and vision models to detect harmful content. It also provides a custom term list feature for customization.
• Azure AI Content Safety has an interactive studio for exploring and testing the service capabilities, while Azure Content Moderator does not.

No prompts or completions are stored for the purposes of content filtering. No prompts or completions are used to train, retrain, or improve the content filtering system without the customer's consent.

To learn more about Microsoft's privacy and security commitments, visit Data, privacy, and security for Azure AI Content Safety.

No. Microsoft hosts the Azure AI Content Safety models within our Azure infrastructure. All customer data sent to Azure AI Content Safety remains within Azure AI Content Safety during processing and in the region you chose and is not transmitted to other regions.