Azure AI Foundry moderation_blocked for safe child images

Andres Hurtado 20 Reputation points
2026-05-07T21:46:32.7733333+00:00

We are experiencing inconsistent moderation behavior when using gpt-image-2 in Azure AI Foundry. Our application generates stylized watercolor children's book characters from uploaded reference images. Some requests fail with: "code": "moderation_blocked" However, retrying the exact same request with the same image and prompt sometimes succeeds. This behavior appears non-deterministic and is affecting production reliability. We would appreciate clarification on: - whether this is expected behavior - best practices to reduce false positives - whether any configuration options exist for this scenario

Foundry Tools
Foundry Tools

Formerly known as Azure AI Services or Azure Cognitive Services is a unified collection of prebuilt AI capabilities within the Microsoft Foundry platform

Answer accepted by question author
Anshika Varshney 15,625 Reputation points Microsoft External Staff Moderator
2026-05-13T03:24:15.6033333+00:00

Hey Andres,

It looks like you're seeing content filtering being triggered for requests that seem safe, especially with gpt-image-2. Let me clarify how this works and what you can check.

  1. Is this expected

Yes, this can happen.

Azure AI content filtering is based on machine learning models that classify content into categories like violence, hate, sexual, and self harm. These models are probabilistic, which means the same input may sometimes pass and sometimes get blocked.

So in some cases, even safe-looking content can be flagged depending on how the model interprets the input.

  1. Why this happens

There are a few common reasons for this kind of behavior:

  • The model may interpret parts of the prompt or image differently each time
  • Certain visual patterns or wording may look ambiguous to the filter
  • The filtering threshold is set to block medium and high severity by default

Because of this, small changes in input can lead to different results.

  1. What you can do to reduce false positives

You can try the following:

  • Check the content filter result returned in the response This helps identify which category triggered the block
  • Slightly adjust or simplify your prompt Avoid any wording that could be interpreted as sensitive or ambiguous
  • Try small changes to the image Cropping, resizing, or changing minor details can sometimes help
  • Test multiple requests This helps identify patterns and common triggers in your scenario
  1. Configuration options

In Azure AI Foundry, content filters are applied by default to all models.

You can:

  • Review your current filter settings under Guardrails and controls
  • Adjust filtering level where allowed
  • For more advanced customization, some options may require special approval depending on your subscription and access

Useful references

I Hope this helps. Do let me know if you have any further queries.

Thankyou!

Was this answer helpful?

2 people found this answer helpful.
0 comments

1 additional answer

Sort by: Most helpful
Most helpful Newest Oldest
  1. kagiyama yutaka 5,490 Reputation points
    2026-05-12T14:51:17.12+00:00

    I think moderation can shift on edge cases, so keeping prompts clearly fictional and unambiguous is the only stable way; there’s no user‑side relaxation, and public docs note filters may still flag context‑dependent edge cases.　

    Was this answer helpful?

    0 comments

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.