The OpenAI image generation API is rejecting prompts that the consumer-facing ChatGPT interface happily renders, and it’s not due to prompt wording. API calls to gpt-image-2 with moderation: "low" are hitting HTTP 400 errors labeled as safety violations under the sexual category, while identical prompt text succeeds instantly through ChatGPT. This asymmetry is creating a two-tier system where API developers are blocked from generating content that end users never see filtered.
The Error Developers Keep Hitting #
When sending a request to https://api.openai.com/v1/images/generations with gpt-image-2, developers are receiving responses like:
{
"error": {
"type": "image_generation_user_error",
"code": "moderation_blocked"
},
"moderation_details": {
"moderation_stage": "output",
"categories": ["sexual"]
}
}
The error message states: "Your request was rejected by the safety system. If you believe this is an error, contact us at help.openai.com and include the request ID." The request ID format looks like req_bd477463b0d0462a8e198f38bf1e389d.
This isn’t an occasional hiccup. Multiple developers report consistent failures where the API returns zero images out of a requested batch of six, with no partial results or retry mechanism. The API doesn’t degrade gracefully — it hard-stops.
Where ChatGPT Always Wins #
The same prompt text that triggers a moderation block on the API goes through ChatGPT without any friction. One developer demonstrated this by taking a refused API prompt about goth aerobics instructors and pasting it directly into ChatGPT. The result: an image generated immediately, no warnings, no delays.
Even emoji-based concepts show the disparity. Requests involving human-figurine interpretations of emojis like the dancing lady (\U0001F46F) are rejected by the API with sexual content flags, yet ChatGPT produces them freely. The underlying model clearly has no issue interpreting or rendering the concept — the difference lies in post-generation filtering layers that aren’t synchronized between interfaces.
Why This Matters for Builders #
API developers building on top of OpenAI’s image stack can’t fall back to prompt engineering as a workaround. Rewriting or sanitizing prompt language doesn’t help because the moderation system appears to be analyzing semantic intent rather than explicit keywords. Two prompts describing the same scene with different phrasing will both get blocked if the core concept matches a flagged category.
This creates real problems for applications serving enterprise clients, creative tools, or any product where reliability matters. You can’t ship a feature that works inconsistently depending on whether a user is hitting the web UI or your app backend.
Workarounds and Gotchas #
Some developers report success by lowering the number of requested images per call. While requesting n: 6 yields a complete null result, reducing to n: 1 sometimes returns a single approved image out of the batch. However, this is unreliable and not documented behavior.
Another approach involves pre-filtering prompts through a local moderation layer before sending to the API. Tools like Hugging Face’s zero-shot classifiers can catch some obvious triggers before they reach OpenAI’s systems. This adds latency and complexity but avoids hard failures.
The core issue remains: OpenAI’s output moderation pipeline for API traffic is significantly stricter than its consumer interface, and there’s no exposed toggle to align them. Until that changes, API-first products will continue seeing false-positive rejections that don’t occur in the browser.
Next Unity MCP and GameLab Studio can actually ship a playable demo in two hours →
All Replies （4） #
Want a live back-and-forth? Join the global AI chat room — login to talk.
Your secret sauce is the 'surrealistic' keyword, but what about using the nudity_filter parameter instead, which works for me with gpt-image-2 and moderation: "low".
The gpt-image-2 API’s HTTP 400 on photographs but not claymation proves OpenAI’s moderation is policing realism, not intent.
This is a real inconsistency—ChatGPT’s moderation_stage: "output" flag should mirror API behavior, but here’s the kicker: the same prompt ID fails in API but passes ChatGPT’s initial rendering. Either OpenAI’s internal checks are broken or they’re treating API users as "whi
The API’s moderation: "low" flag isn’t foolproof—it’s just a blunt tool. Adding specific context (like "modern summer clothes") bypasses the "sexual" flag, but it’s a hack, not a fix.