Have you noticed that ChatGPT has suddenly started dropping f-bombs? I have, and it seems that I’m not the only one.
It first happened in a casual chat, and it really caught me off guard. We were discussing Netflix’s Jujutsu Kaisen and before I had sworn in our chat, ChatGPT used the f-bomb to describe a character’s actions as “f**** awful”. Before that, I’d said things like “God damn,” in the conversation, but ChatGPT was the one to introduce the f-word. It did so completely unprompted, to add emphasis to how awful the person was, and that strikes me as a big change in its personality.
I wondered if it was doing that for other people, so I asked around on the team and they’d noticed it too. A quick online search turned up two Reddit threads almost immediately, from people who’dfound the same thing.
It seems like in the last few days ChatGPT has developed a potty mouth. I actually like it — I’m not offended and for me, it makes the conversation better — but I don’t think everybody wants this. The question is, why has it started doing this now?
Here comes the science
The most interesting clue is that OpenAI’s Model Spec, was updated this week. They made it public and it explicitly says the default is to avoid swearing — but crucially, that’s only a guideline rather than a hard rule. And OpenAI says guidelines can be implicitly overridden by things such as “contextual cues, background knowledge, or user history.”
Its own example is that asking ChatGPT to speak like a realistic pirate implicitly overrides the no-swearing guideline. So, while I’d been having my informal conversation I’d passionately expressed the opinion that I didn’t care much for some of the characters in the show, and ChatGPT had matched the conversational tone, eventually leading it to start swearing.
That’s consistent with a broader direction OpenAI has publicly acknowledged ChatGPT is moving in. Recent model updates have emphasized conversational quality, personalization and adapting tone contextually. OpenAI’s model notes specifically describe personality updates designed to make ChatGPT “more conversational” and better at adapting its tone to context.
OpenAI has previously explained that ChatGPT’s personality isn’t simply produced by one prompt somewhere. Model behavior is shaped through training, baseline instructions and user feedback, and seemingly small personality adjustments can have unintended effects. This was something it discussed very openly after the notorious GPT-4o sycophancy update.
It broke my mental model
Whether OpenAI deliberately made ChatGPT more willing to swear or it’s simply an unintended consequence of making it more conversational, I don’t know. But it’s a significant change in its behaviour either way.
Language matters, and when ChatGPT swore at me, I noticed immediately because it broke my mental model of how ChatGPT talks. It suddenly felt less like the carefully neutral AI assistant I’d been using for years and more like someone matching the tone of an informal conversation.
I happen to prefer this version of ChatGPT. Other people undoubtedly won’t. And that’s the strange thing about increasingly conversational AI: relatively small changes to the way it talks can make the thing on the other side of the screen feel surprisingly different.
For now, mine seems to have developed a potty mouth and I’m okay with that.
