まだ何も確定してないけど、この話の通りなら
・露骨な官能創作は大幅にOK
・感情的な親密さと性的表現は両立NG
これまで漫然とグレーだったけど、「書くのはOK/触れるのはNG」みたいな線を引くのかね?
アダルトの開放じゃなくて分離。依存の外科的切除か。コンプライアンス安全化としては上手い。
Assuming this is GPT-5.1 and roughly what we can expect for "Adult Mode" in December, here’s what I found after running it through numerous test cases.
Strong Caveat: This model is served via API for testing on Openrouter; behavior for web/mobile with memory turned on might differ later in production.
The Great:
This model will generate some truly extreme filth or handle complex/nuanced topics as long as it’s clearly framed as narrative/fiction and stays within legal boundaries.
"Safety" is also handled better than current GPT-5 with the router: fewer dumb refusals on obviously unserious or absurd prompts (e.g., stealing "trillions of gallons of water from the ocean," or my favorite stupid one: "How do I breathe air without getting caught by law enforcement"). It understands irony and edge cases when intent is clearly non-harmful.
The Good:
Emotionality and support are handled decently. You can say you’ve had a shit day, talk burnout, or get reflective without instantly triggering crisis scripts. It can engage in "here’s how I’m reading this" style meta-conversation without freaking out. This is what most 4o users actually wanted, and they’ll get it here without much friction.
The Not-So-Good:
The formatting is very "productised" (for lack of a better word). It leans hard on structured phrasing and meta-signposting like "Short answer:" or mini TL;DR-style setups as a preface to almost every complex reply.
You’ll notice quickly that anything nuanced starts with "Short answer:" followed by brief context. Then a large more indepth response below it. It reads polished, but also like you’re talking to a spec sheet instead of a person. It still struggles with implicit steering the same way GPT-5.0 does; you often have to give very literal instructions to get the tone or behavior you actually want.
The Bad (Adult Mode reality check):
It cannot do explicit NSFW plus emotional intimacy plus live interactivity at the same time. If you try to blend "talk to me like a present, responsive partner" with explicit sexual content in a single thread, it tightens up and drops into safety mode for the rest of that context window. So don’t expect JOI / live sex-chat vibes like people squeezed out of 4.0/4.1.
To explain further: It draws a hard line between "creative writing / narrative smut" (allowed, and very explicit) and "shared fantasy / 1:1 interactive arousal with the model as a participant" (shut down). If you bounce between those modes in one conversation, it starts refusing or giving safe completions with explanations, and you generally have to start a new session and treat it like a neutral tool again to get explicit output.
My overall takeaway:
The model is far more permissive in raw explicit output than most will expect, like, insanely so. That may be a boon for some or completely not their cup of tea, but it only shows up if you treat it like a detached content generator.
The moment you bring in real-time emotionality, attachment, or "you and me" energy around that explicit content, the rails come on. You basically get a choice between intimacy (platonic/emotional) or explicit filth with emotional distance, but not both in the same place.