Content filters hitting with no explanation despite being set to annotate only

Hello Graham LaBossiere,

Greetings! Thanks for raising this question in Q&A forum.

This is a great observation and you're right to flag it. Here's what's likely happening: Azure OpenAI has two separate layers of content filtering. One is your configurable content filter policy (which you've set to "Annotate only"), and the other is a non-configurable, built-in safety layer that runs independently underneath it. Even simple-looking prompts like "What does <company> do?" can occasionally trip the built-in "Protected material for text" or "Indirect prompt injection" checks depending on how the company name is interpreted internally. The key frustration here — and you've spotted it correctly is that when the block comes from this lower-level system, content_filters comes back as null, giving you no visibility into which filter fired. This does appear to be a transparency gap in the API response.

Here's what you can do step by step to investigate and work around this:

Step 1: Double-check your content filter policy is correctly applied

In Azure AI Foundry (formerly Azure OpenAI Studio), go to Content Filters → confirm your custom policy with "Protected material for text" set to "Annotate only" is actually assigned to your deployment, not just created. A policy that's created but not linked to the deployment won't take effect.

Step 2: Check the full API response for prompt_filter_results

Look at the raw API response for the prompt_filter_results field, not just content_filters. Sometimes the block is triggered at the prompt level (input) rather than the completion level (output), and the details appear there instead.

Step 3: Test with a simplified prompt

Try sending a very neutral, minimal prompt like "Tell me about Microsoft" to see if the same block occurs. This helps isolate whether the issue is the company name itself or something else in your request payload like the system prompt or conversation history.

Step 4: Review your system prompt

If you have a system prompt configured, try temporarily removing it and retesting. Occasionally a system prompt can interact with the built-in filters in unexpected ways, especially if it contains instructions about external entities or company data.

Step 5: Enable diagnostic logging

In your Azure OpenAI resource, turn on Diagnostic Settings and send logs to a Log Analytics workspace. The AzureOpenAIRequestLogs table can sometimes surface filter reason details that are not visible in the API response itself.

Step 6: If the issue persists, raise a support ticket

If after all the above the content_filters field still returns null with no explanation, this is likely a platform-side bug in filter result reporting. Open an Azure Support request with your resource name, deployment name, a sample prompt that reproduces it, and the full raw API response. The Azure AI team can inspect the filter pipeline on the backend for your specific resource.

As a quick note: the null response for content_filters when a block occurs is a known transparency limitation in certain edge cases, and Microsoft has been gradually improving this. Your feedback through the support channel will help push that forward.

If this answer helps you kindly accept the answer which will help others who have similar questions.

Best Regards,

Jerald Felix.