Json parsing error with content moderation

[mdpead]

Confirm this is an issue with the Python library and not an underlying OpenAI API

• This is an issue with the Python library

Describe the bug

Hello,

I've seen an example where I'm submitting to the responses.parse endpoint, and I get back a pydantic validation error. Upon inspection, it looks like the model is returning "I'm sorry, but I cannot assist you with that request.", which obviously fails to json parsing step.

I'd expect it to raise something like a APIResponseValidationError (if not something more specifically related to moderation), instead of a pydantic_core.ValidationError.

To Reproduce

I've found it very difficult to reproduce as the original example is from internal business documents I cannot share, and my attempts to trigger that response have failed.

The basic steps are:

1. Create client
2. Submit problematic prompt to client.responses.parse with text_format= pydantic schema
3. Observe error

For context, I am using the async client, through Microsoft AI Foundry, with chatgpt-4.1-mini, and the content filter is DefaultV2. The Foundry API version is 2025-03-01-preview.

Code snippets

from openai import AsyncAzureOpenAI
credential = DefaultAzureCredential(logging_enable=True)
openai_client = AsyncAzureOpenAI(
        api_version="2025-03-01-preview",
        azure_endpoint=xxx,
        azure_ad_token_provider=get_bearer_token_provider(
            credential,
            "https://cognitiveservices.azure.com/.default",
        ),
)
await openai_client.responses.parse(
    input=prompt,
    text_format=pydantic_schema
)

OS

Windows 11

Python version

Python 3.12.3

Library version

openai 2.15.0

[CodeVishal-17]

Thanks for the detailed report — this does look like a real SDK-level issue.

responses.parse() appears to assume the model will always return JSON-conformant output, but when content moderation is triggered the model instead returns a plain-text refusal. That refusal then falls through to Pydantic validation, resulting in a low-level pydantic_core.ValidationError.

From a consumer perspective, this feels like an error-handling gap in the Python client. Moderation-triggered responses are expected runtime behavior and should likely be surfaced as a higher-level SDK exception (e.g. a moderation-specific error or a wrapped APIResponseValidationError), rather than leaking raw schema validation errors.

Even if the underlying API response is valid, responses.parse() probably needs a guard to detect non-JSON / refusal responses before attempting Pydantic validation.

Happy to help test a fix or provide a minimal repro if useful.