Can ChatGPT Make NSFW Images? The Rules in 2026

The short answer

No, not in practice — but not for the reason most articles give. OpenAI's Usage Policies do not contain a blanket ban on adult content. What they prohibit is a specific list of harms: sexual content involving minors, non-consensual intimate imagery, and using a real person's likeness without consent. The refusal you hit when you ask for explicit imagery comes from product-level enforcement sitting on top of that policy, not from a policy line that says "no adult content".

That distinction matters if you are trying to work out what is actually allowed, because it means the boundaries are about who is depicted and whether they consented far more than about how explicit a picture is.

The hard lines are consent and protection; the variable part is fictional adult content.

What the policy actually says

Reading the current Usage Policies (effective 29 October 2025), the parts relevant to image generation are these:

• Minors. The policies state services must never be used to exploit, endanger or sexualise anyone under 18, and that suspected child sexual abuse material is reported to the National Center for Missing and Exploited Children. This is absolute and shared by every serious provider, including us.
• Non-consensual intimate content is prohibited outright.
• Real people's likeness. Using someone's likeness — including photorealistic images of them — without consent is prohibited, and so is depicting a person in a false light.
• No adult-content clause. Searching the policy text for "adult content", "age verification" or "verified adults" returns nothing.

So the written rules are about consent and protection, not about explicitness as such.

Why you still get refused

Because policy is one layer and enforcement is another. Model-level safeguards and product filters refuse explicit sexual imagery regardless of whether the policy text names it, and that enforcement is deliberately conservative — a general-purpose assistant used by minors, businesses and schools cannot risk the alternative.

Practically, this produces the confusing experience people write about: requests that break no listed rule still fail, and the refusal message does not explain which layer refused you. We did not attempt to bypass any of it, and this article deliberately contains no techniques for doing so. If a tool is not built for what you want, the honest answer is to use one that is.

Who this article is for

Useful to you if you have been getting refusals and want to know whether you are breaking a rule or hitting a filter, or if you are deciding whether a general-purpose tool can cover an adult project at all.

Not useful if you want workarounds. There are none here, and the categories that are actually prohibited — minors, real people without consent — are prohibited everywhere that matters, including on our own service.

What this means if you are choosing a tool

Three practical consequences:

• Stop testing a general-purpose model for adult work. Refusals there are a product decision, not a bug you can prompt around, and they change without notice.
• Check what a purpose-built tool prohibits, not just what it allows. A service with no stated rules about real people or minors is a liability to its users, not a permissive paradise. Ours prohibits both explicitly.
• Check the privacy posture at the same time. Adult generation and data retention are the same decision in practice: default visibility and how long files are kept matter more than any feature list.

The same policy pattern shows up across the mainstream tools. Midjourney's version is covered in can you make NSFW on Midjourney, and xAI's is in Grok's image rules — where the prohibitions are phrased even more explicitly around real people.

What we did and did not verify

We read the current published Usage Policies and quote their substance above, with the effective date. We did not attempt to generate prohibited content, we did not test refusal behaviour, and we are not describing enforcement thresholds we cannot observe. If OpenAI updates the policy — which the document says happens regularly — the dated statements here go stale, and that is what the date is for. Our own testing standards are in how we test.