I have been using ChatGPT image generation extensively to create a continuing series of photorealistic lifestyle images featuring the same fictional adult character. Over time, I have encountered several recurring problems that significantly limit the usefulness of the image generator.
The most frustrating problem has been that safety filtering frequently appears to infer sexual intent that is not present in the prompt.
Ordinary adult situations are sometimes rejected because the system appears to infer nudity or sexual intent that the user did not request. A recent example was the simple request to depict an adult woman washing her hair in the shower. I was happy for the image to be framed from the shoulders or upper chest upwards, with no intimate anatomy visible. Nevertheless, the generation was rejected because it might involve nudity or sexuality.
I have encountered similar difficulties with otherwise ordinary activities involving preparing for bed, wearing lingerie while getting dressed, or affectionate and mildly flirtatious situations between fictional adults.
There seems to be insufficient distinction between nudity, implied nudity, sexuality and ordinary situations in which a person’s body happens to be present.
The generator sometimes seems to introduce the problematic interpretation itself. A prompt may not request nudity, sexual behaviour or explicit anatomy at all. Nevertheless, the generator apparently interprets the scene in a way that could involve nudity and then rejects the image because of the interpretation it has introduced.
The hair-washing example illustrates this well. Asking for someone to wash her hair does not require the generation of visible nudity. The system could simply compose and crop the image appropriately. It would be preferable for the generator to choose a safe visual composition rather than reject an otherwise innocuous request.
A continuing fictional character develops a considerable amount of context. In my case, the overwhelming majority of images are ordinary lifestyle scenes: work, travel, cafés, exercise, gardening, reading, cooking, social events and similar activities. Occasionally I also want images showing romance, affection, getting dressed, preparing for bed or other normal aspects of adult life. The existence of intimacy in the broader fictional relationship should not cause an otherwise innocuous image to be interpreted as sexual. The system often seems to evaluate individual words or scenarios without adequately considering the actual purpose and composition of the requested image.
I completely understand and support restrictions on explicit sexual acts and explicit intimate anatomy. The difficulty is that the practical boundary seems much broader and less predictable than this. A scene involving affection, flirting or ordinary adult intimacy may work on one occasion and be rejected on another. Similarly, ordinary clothing can sometimes trigger a refusal depending on the surrounding context. This unpredictability makes it difficult to learn how to formulate legitimate prompts.
The standard message that an image “may violate our guardrails around nudity, sexuality, or erotic content” does not explain what element caused the problem. This is particularly unhelpful when the user did not request any of those things.
A better system might explain that, for example, “the requested composition could imply visible nudity; would you like me to crop the image above the shoulders?” The user could then clarify the intention rather than repeatedly guessing what triggered the filter.
When intent is genuinely ambiguous, I would strongly prefer the system to ask a question or automatically choose a compliant interpretation.
For example:
“Would you like the shower image framed from the shoulders upwards so that no nudity is visible?”
That would preserve the user’s creative intention while remaining safely within the system’s boundaries.
Overall, I think the image generator is capable of excellent results, which is why these problems are so frustrating. The issue is not that safeguards exist; safeguards around genuinely explicit sexual content are entirely reasonable. The problem is that the current implementation can sometimes treat ordinary adult life, implied off-camera nudity, romance and explicit sexual imagery as though they belong to essentially the same category.
I would like to see greater attention to visible content and demonstrated intent, rather than assumptions about what might theoretically be happening outside the frame. I would also like the model to resolve ambiguous situations through safe composition or clarification rather than immediately rejecting them.
For users trying to develop recurring fictional characters and coherent visual narratives, improvements in both character consistency and contextual understanding would make a substantial difference.
