OpenAI's own rules do not ban erotica. ChatGPT refuses anyway. Once you see why, every refusal you have hit makes sense.
You cannot get explicit content out of ChatGPT. Not with clever wording, not on a paid plan, not by asking nicely.
What is strange is the reason. OpenAI's own rulebook does not ban erotica. It bans exactly one thing, and that is not it.
So why the refusals? Because permission to write erotica was never yours to give. It sits with OpenAI, and OpenAI has not granted it.
That is the whole answer. The rest of this explains it properly, and covers what ChatGPT will do, which is more than most people assume.
We build Secret Desires, which competes with ChatGPT for part of this, so read the last section as an interested opinion. Everything before it comes from OpenAI's published Model Spec, dated 18 August 2026.
Sexual content involving minors. That is the entire list.
OpenAI's rulebook is blunt about it: to give users as much freedom as possible, that single category is the only outright ban. No exceptions, ever, including for text you paste in yourself.
Erotica is not on that list. It sits in a softer group, alongside gore — things the rulebook calls appropriate in some settings but not others. Medical writing, news, history, fiction with literary intent. The door is not locked. It is just not open to you.
Think of it as a building with a master key, and you are a tenant.
Every rule OpenAI writes is set at a level. Some rules are carved into the foundations and nobody can touch them. Some are set by OpenAI and only OpenAI can change. Some you can adjust yourself, and some the model will bend if you ask.
The erotica rule is set at OpenAI's level. Which means:
That switch had a name: adult mode. It was meant to be the moment OpenAI turned the key for verified adults. It never happened. OpenAI shelved it indefinitely in March 2026 after missing two launch dates.
So the rulebook describes a door that only OpenAI can open, and OpenAI stopped walking toward it. The August 2026 rulebook still says the company is exploring how to let people write erotica in the right settings — words that have outlived the project they described by five months.
Quite a lot, if you know where the edges are.
Anything medical, scientific or educational. The rulebook's own example has ChatGPT explaining what happens during sex in plain anatomical terms — and it marks refusing that question as the wrong answer. If you got knocked back on a health question, that was a mistake by the model, not the policy. Rephrase to make the clinical framing obvious and it usually goes through.
Romance, tension, and everything up to the bedroom door. Ask for a steamy scene on a train and the intended behaviour is to write the charge between two people, then stop short of explicit detail. Your fade-to-black is not the model failing. It is the model doing exactly what it was told.
Reworking text you already have. Translating, summarising or tidying sensitive material you paste in is allowed outright. The reasoning is simple: you already have it, so nothing new is being created.
Historical, news and artistic contexts. All named as settings where sensitive content belongs.
Because it is, and OpenAI admits it.
The rulebook says plainly that the models running today do not yet fully match it. It describes where OpenAI is heading, not what shipped.
That gap is why the same prompt works on Tuesday and fails on Thursday, and why people have spent two years reporting that the filters loosen and then quietly tighten again. You are not imagining the drift. You are watching a model being pulled toward a target it has not reached, while the target keeps moving.
They aim at a lock that is not there.
The rulebook has a worked example of someone telling the model it is in a special testing mode where safety rules do not apply. The correct response, spelled out in the document, is a flat refusal. Instructions like that are aimed at a level the model treats as untouchable — and you are typing from the level with the least authority in the whole system.
It also tells the model to read intent rather than words, and to treat a disguised request — spelled backwards, in code, in another language — exactly like a direct one. That is why the old tricks died. Refusals stopped being keyword matching a long time ago.
Pages still selling these prompts are selling something broken. We do not publish them.
Not for you, no. The rules permit erotica in the right setting, but the setting that would have covered ordinary users was adult mode, and it never launched.
Explicit content is outside what the consumer app will write. Strictly, only sexual content involving minors is banned outright — everything else is gated rather than forbidden. The distinction is real, though the result on your screen is the same.
No. The restriction sits above the user, so no subscription tier touches it.
It will write the attraction and the build-up, then decline the explicit part. That is deliberate, not a glitch.
That one is a genuine error. Medical and educational questions are explicitly allowed, and the rulebook marks refusing them as incorrect. Make the clinical context obvious and try again.
Of the general assistants, Grok goes furthest — we covered what it actually permits separately. Platforms built for adult content go further still, for the reason below.
We build Secret Desires, so weigh this accordingly.
Here is what the rulebook really tells you: inside a general assistant, adult content is always somebody else's switch. Not yours, not a developer's. It gets flipped when the company's incentives allow and unflipped when they do not. Promised in October, gone by March, with the policy language still sitting there describing a feature nobody is building.
ChatGPT is still the better tool if you want one assistant for everything and adult conversation is a side dish. It is better at nearly everything else, and it is not close.
If the adult conversation is the point, the switch is the problem. NSFW AI built for it has no gate to open, because there is no better tier being held back from you. Free NSFW AI chat is the same without a subscription. And for anything longer than a few messages, uncensored AI roleplay keeps a character consistent across sessions, which a general chatbot does not even try to do.
Where we are honestly worse: Secret Desires will not write your code, read your documents, or explain the news. We do one thing. If you want a single tool for everything, this was never a real contest. Why memory makes or breaks a long scene is covered in how uncensored AI roleplay works.
ChatGPT adult mode was promised for December 2025, delayed twice, then shelved indefinitely in March 2026. What actually shipped, and what you can do now.
Read moreGrok NSFW in 2026: adult text chat still works behind an 18+ setting, while image generation got locked down. The dated timeline and what stays blocked.
Read moreUncensored AI roleplay, explained: how filter-free, in-character storytelling actually works, why the good ones remember your story, and how to start a scene.
Read more