ChatGPT refuses an image when one of three safety layers flags it: the chat model itself declines, a classifier blocks the prompt or uploaded photo, or a second check blocks the finished image. The usual triggers are sexual or suggestive content, anything involving a real person's likeness, photos that look like they include a minor, graphic violence, and "in the style of" a living artist. Many refusals are false positives, and a clearer prompt often fixes them.

Disclosure: VdoBloom is our platform.

The message "I can't create that image" or "this request may violate our content policies" rarely says which rule you hit. That is frustrating when the picture you asked for is harmless. This guide explains what OpenAI's image rules actually target (as of September 2026), the most common reasons a normal request gets caught, how to reword a request without trying to trick anything, and when it makes more sense to use a different image model.

The three places ChatGPT can stop an image

OpenAI's own system card for ChatGPT's native image generation describes a layered setup. Knowing the layers helps you guess what went wrong:

Chat model refusal. The assistant reads your whole conversation, not just the last message, and can decide not to call the image tool at all. If you get a polite paragraph instead of an image, this layer made the call.
Prompt blocking. After the image tool is called, text and image classifiers check the final prompt and any uploaded photos. A flag here usually produces a short "content policy" message.
Output blocking. The image is generated, then a separate check (including child-safety classifiers and a reasoning monitor) reviews it. This is why an image can appear to start rendering and then disappear, or why the same prompt works once and fails the next time.

Because the third layer judges the actual pixels, a harmless prompt can still fail if the model happened to render something the checker reads as unsafe, such as too much skin or blood. Running the same request again sometimes works for that reason.

What the ChatGPT content policy for images actually targets

OpenAI's usage policies and its published image-generation safety notes point to a fairly consistent list. As of September 2026 the main categories are:

Sexual and erotic imagery. OpenAI says it aims to prevent erotic or sexually exploitative images, with extra safeguards against non-consensual intimate imagery and sexual deepfakes of any kind.
Minors. Any sexualization of anyone under 18 is banned outright. OpenAI also said at launch that editing uploaded photos of photorealistic children would not be allowed, and it runs a classifier on every uploaded photo to predict whether it shows a child.
Real people. OpenAI's policies prohibit using someone's likeness without consent in ways that could mislead. Public figures can be depicted in some contexts (satire, commentary), but not in erotic, violent or hateful material, and public figures can ask to opt out.
Graphic violence and self-harm. Violence in fiction and art is generally allowed; photorealistic, gory violence in certain contexts is not. Anything that promotes self-harm is blocked.
Hate and extremism. Propaganda or recruitment imagery is blocked; hateful symbols can appear in neutral, educational or critical context.
Living artists. OpenAI added a refusal when a prompt asks for an image "in the style of" a living artist.
Deception and fraud. Fake IDs, documents, and images meant to pass as real evidence fall under the broader usage policies.

OpenAI changes these rules from time to time, so check the current OpenAI usage policies if a refusal seems to contradict what you read here.

Why ChatGPT won't generate images of people (the common false positives)

Most frustrated searches are about requests that are not against any rule. These are the patterns that trip the filters most often:

Adults who look young

OpenAI states that its photo classifier is tuned to err on the side of caution and that "younger-looking adults may be incorrectly flagged as children." If you upload a selfie and ask for an edit, a youthful face, a school-style uniform, or a small frame can be enough to block the edit, even though you are an adult.

A child somewhere in the frame

If a photo contains both an adult and a child, the classifier is designed to label it as a child image. A family photo, a kid in the background at a wedding, or a poster of a child on a wall can block an edit to the adult in front.

Swimwear, gym and fashion shots

Beach photos, fitness progress shots and lingerie product listings are legal and common, but they sit near the sexual-content line. Words like "sexy", "hot" or "seductive" push the chat model toward refusing, and the output check can block a result that shows more skin than you intended.

Medical, historical or fictional violence

A wound for a first-aid poster, a battle scene for a history lesson, or a horror book cover can read as graphic violence if the prompt focuses on blood and injury instead of the context.

Leftover context from earlier in the chat

The first layer reads the whole thread. If an earlier message was borderline, later innocent requests in the same conversation can keep getting refused. Starting a new chat clears that context.

Named artists, brands and characters

Asking for a living illustrator's style, a trademarked character or a real logo can be refused for intellectual-property reasons, not safety ones.

How to reword a refused request (legitimately)

Rewording is appropriate when your request is allowed and the filter misread it. It is not a way to get content the policy prohibits, and trying to disguise prohibited content can get an account restricted. With that in mind:

Say the purpose. "A product photo of a one-piece swimsuit for a shop listing, adult model, neutral studio pose" gives the chat model context that "hot girl in a swimsuit" does not.
Describe the look, not a person's name. Instead of a living artist, describe the medium: "loose watercolor, muted earth tones, visible paper texture, ink outlines."
Replace loaded adjectives. "Confident", "elegant", "editorial" and "athletic" describe a style without implying sexual content.
Crop the upload. Remove children, ID cards, screens, and anyone who did not agree to be in the edit before you upload.
Start a new chat if a thread keeps refusing unrelated requests.
Ask what triggered it. ChatGPT will often tell you which part of the request it objected to, and you can decide whether that part matters.

If the honest version of what you want is sexual content of a real person, an edit of a child, or a deceptive image of someone, no rewording should get it through, and none of the tools below will make it either.

"ChatGPT won't generate images anymore": not always a policy issue

Sometimes the image tool fails for reasons unrelated to content. Free accounts have tighter image limits than paid plans, peak hours cause errors and slowdowns, and outages happen. If simple prompts like "a red apple on a table" also fail, check OpenAI's status page and your plan's limits before rewriting your prompt. A content refusal usually says "policy"; a limit or outage usually says to try again later.

When a different image model is the better fit

ChatGPT's filters are tuned for a very broad audience, and they are stricter than some other models on things like fashion, swimwear and stylized fantasy art. On VdoBloom's Text to Image tab you can pick the model per image. Each model shows a content badge: STRICT models apply their provider's own filters, while HOT models have fewer restrictions on clothed, adult, non-explicit content. Prices below are per image at the default resolution:

Model	Badge	Credits per image
GPT Image 2 (same OpenAI model family)	STRICT	3 at 1K, 5 at 2K, 8 at 4K
Qwen Image 2.1	HOT	3 at 1K, 5 at 2K
Qwen Image 3.0	HOT	5
Seedream 4.5	HOT	7
Seedream 5.0 Pro	HOT	10 (Standard), 19 (High)
Nano Banana Pro	STRICT	18 (24 at 4K)

Two honest notes. First, choosing GPT Image 2 inside VdoBloom does not remove OpenAI's filters; it is still a STRICT model. If ChatGPT refused on content grounds, a HOT model such as Seedream 4.5 or Qwen Image 2.1 is the relevant switch. Second, "fewer restrictions" is not "no restrictions." VdoBloom runs its own check before charging credits and blocks nudity and undressing requests, and it always blocks any sexualized content involving minors. If a provider's filter blocks a generation after it starts, the credits are refunded.

A worked example: ten 1K test images on Qwen Image 2.1 cost 30 credits; the same ten on Seedream 4.5 cost 70. New accounts get 10 free credits, enough for three Qwen Image 2.1 images at 1K. For a comparison of permissive models, see the best uncensored AI image generators.

Consent and real people

Most image refusals involving people exist for a real reason, and the rules on VdoBloom are the same in spirit:

Only upload your own photos or photos of adults who have agreed to the edit. VdoBloom's upload tabs require you to confirm that everyone pictured is 18 or older and that you have the rights and consent to use the image.
Never upload or generate images of minors in any suggestive context. VdoBloom blocks it.
Never create intimate or sexual images of a real person. That is non-consensual intimate imagery and is illegal in many countries, regardless of the tool.
Do not use a public figure's likeness in anything meant to mislead people about what they said or did.

Frequently asked questions

Why won't ChatGPT generate images of people?

It does generate people in most cases. Refusals involving people usually come from an uploaded photo the classifier reads as a minor, a request that looks sexual or violent, or a real person's likeness used in a way that could mislead. A new chat with a clearer, purpose-first prompt fixes many false positives.

What does "this request may violate our content policies" mean in ChatGPT?

A classifier flagged the prompt, the uploaded image, or the generated result. It does not tell you which category. Check your request against the list above: sexual content, minors, real people, graphic violence, hate, living-artist styles, and deceptive images.

Why does ChatGPT refuse the same prompt that worked yesterday?

The final check reviews the generated image, which is different every run, and the chat model considers the whole conversation. A slightly different result or different earlier messages can change the outcome. OpenAI also updates its safety systems over time.

How do I make AI images of myself with ChatGPT?

Upload a clear, recent photo of yourself alone, crop out other people, and describe the change plainly, for example "same person, business headshot, gray backdrop." If it keeps misreading your photo as a minor, a dedicated tool such as AI Headshots or an image-edit model may work better.

Is there a way to bypass the ChatGPT image filter?

We do not recommend trying, and it can get your account restricted. If your request is allowed, add context and remove ambiguous words. If it is content ChatGPT does not allow but the law and consent do, use a model whose policy permits it rather than trying to trick one that does not.

Why ChatGPT Won't Generate Your Image: Real Reasons