Share via

Azure OpenAI returns "I cannot assist" despite content filters set to Lowest blocking - false positives on legitimate business prompts

Mujtaba 5 Reputation points
2026-04-14T12:18:30.81+00:00

My Azure OpenAI resource is returning "I'm sorry, but I cannot assist with that request" for legitimate business requests (medical appointment booking and job portal features and many more simple requests). I have already configured both Input and Output content filters to "Lowest blocking" (High severity only) for Violence and Self-harm categories. The issue persists even for benign prompts. I need Microsoft to investigate this false positive and unblock my use case.

Azure OpenAI in Foundry Models

3 answers

Sort by: Most helpful
  1. Ghulam Muhayyu Din 0 Reputation points
    2026-04-20T10:09:43.7433333+00:00

    Hello Mujtaba,

    Thank you for sharing your workaround! That is a fantastic observation regarding the regional behavioral differences with the gpt-5.4-mini deployment.

    Your hypothesis about server load in eastus2 is highly plausible. In cloud environments, we sometimes observe that regional capacity constraints or staggered rollouts of backend safety updates can cause the system to default to a generic fallback refusal (like "I cannot assist") rather than returning a standard rate limit or capacity error.

    I am glad to hear that migrating your deployment to the eastus region completely resolved the false positives for your business prompts.

    I highly recommend posting your region-switch finding as a new, separate answer on this thread and clicking "Accept Answer" on it. This will mark the thread as solved and be incredibly helpful for other developers who might be encountering the same localized issue in eastus2!

    Was this answer helpful?

    0 comments No comments

  2. Manas Mohanty 16,840 Reputation points Microsoft External Staff Moderator
    2026-04-20T01:41:29.9266667+00:00

    Hi Mujtaba,

    Good day.

    Agreed to below community member suggestion.

    Yes, it would be recommended to opt for Modified Guard rails access which would disable content filtering and avoid false positive (Considering Production environment)

    (Currently limited to enterprise customer, approved from Gating team based on use case, please provide org email and justification on above form.)

    Overall recommendation

    1. Please test GPT 5.1 or 5.2, Restriction seems to be bit higher in GPT 5.4 models
    2. Create a new content filter with lowest threshold to isolate any regression issue
    3. opt for Modified content filtering (Guard rails) access
    4. Please share the prompts and model used in private message if you are still looking for investigation on "False positive content filtering"

    Sorry for the inconvenience.

    Thank you for your inputs.

    Was this answer helpful?


  3. Ghulam Muhayyu Din 0 Reputation points
    2026-04-19T17:24:04.6633333+00:00

    Hello Mujtaba,

    Dealing with false positives is a frequent challenge when analyzing model robustness an issue I often tackle in my own NLP and machine learning research.

    In Azure OpenAI, the "I cannot assist" message can trigger independently of your configurable content filter policy. The models have a built-in, non-configurable safety layer that can aggressively block business or medical prompts. First, navigate to your content filter policy and disable "Prompt Shields" (specifically Jailbreak and Indirect Prompt Attack detection), which are notorious for over-triggering on legitimate requests. If the false positives persist, you will need to apply for modified content filtering access through Microsoft's official form (aka.ms/oai/modifiedaccess) to grant deployment-level exemptions for your specific business use case.

    Was this answer helpful?


Your answer

Answers can be marked as 'Accepted' by the question author and 'Recommended' by moderators, which helps users know the answer solved the author's problem.