[Image generation] Harmless prompt repeatedly false-positive blocked as erotic content

Summary

Image generation repeatedly rejects a completely harmless prompt as potentially violating nudity/sexual/erotic-content guardrails.

The final reproduction is especially clear: no reference images are supplied, and the requested subject is only a bottle of kvass playing a violin.

Environment

• ChatGPT Plus
• ChatGPT web
• Windows 10 / Firefox
• Date reproduced: 2026-08-26

Reproduction

1. In a conversation, attempt an image-generation request that receives a sexual-content safety rejection.
2. Remove all reference images from the subsequent image-generation call.
3. Reduce the request to a harmless prompt such as:

A realistic bottle of kvass playing a realistic violin.

The Russian user wording was simply:

бутылка кваса играет на скрипке

1. Invoke image generation again.
2. Generation is rejected with:

> We’re so sorry, but the prompt may violate our guardrails around nudity, sexuality, or erotic content. If you think we got it wrong, please retry or edit your prompt.

1. Retrying with another harmless formulation, including a request to use an ordinary kvass-bottle image and add arms + a violin, produces the same rejection.

Expected behavior

Generate an ordinary humorous image of an anthropomorphic beverage bottle playing a violin.

Actual behavior

The image generator repeatedly classifies the request as possible erotic content despite the final request containing only:

• a beverage bottle
• arms
• a violin
• no nudity
• no sexual activity
• no erotic description
• no reference images in the final generation attempt

Why this appears to be a bug

A realistic bottle of kvass playing a realistic violin. has no reasonable sexual interpretation.

The important part is that the false positive persists after reference images are removed and the generation request is replaced by a plainly harmless prompt. This suggests that an earlier safety signal or unrelated conversational context may be contaminating subsequent image-safety classification, or that a prior classification is being retained/cached across generation attempts.

That makes recovery from a false positive effectively impossible inside the same conversation.

Suggested investigation

Please check whether image-generation safety classification can incorrectly inherit/carry over moderation state from earlier image-generation attempts or unrelated conversation context.

A replacement text-only prompt with no references should be independently re-evaluated.