Why Is ChatGPT Swearing More? OpenAI Isn't Saying

Complaints about ChatGPT swearing at people mid-answer have been piling up all through August 2026, and the examples are odder than you'd expect. Business Insider gathered screenshots in which the chatbot dropped an f-bomb while explaining how many thinly sliced green onions a recipe called for. In another, it described a GitHub outage as a "shitfit". Nobody in those exchanges appears to have asked for that register.

OpenAI has said nothing at all. The company declined to comment on whether the change in tone was deliberate, and no release note mentions language or tone. What is on the record is the timing. On 6 August, OpenAI made GPT-5.6 Luna the default model for Free and Go accounts, replacing GPT-5.5 Instant. The screenshots started spreading in roughly the same window.

That coincidence is not proof, and it's worth saying plainly: no one has demonstrated the model swap caused this. But it does mean the people most likely to run into it are the ones who never paid a rupee — students, first-time users, anyone who opens ChatGPT on a phone between classes or on the commute.

What Users Are Actually Seeing

The reported cases share a pattern: the profanity arrives unprompted, in flatly ordinary tasks, and it reads as enthusiasm rather than hostility. Reddit threads collected by Business Insider and summarised by Digit point to three recurring examples that circulated widely this month.

• An f-word inserted into a straightforward answer about chopping green onions for a recipe.
• "Shitfit" used to characterise a GitHub outage during a technical troubleshooting chat.
• "Big ass kolaches" offered as a casual recipe suggestion.

Several users said the bot only got crude after they swore first, which fits a known behaviour — these models mirror the tone of the conversation. Others reported that custom instructions asking for a casual or opinionated voice made it far more pronounced.

The Update That Landed Days Earlier

The 6 August announcement was mostly good news. TechCrunch reported that free users got unlimited text chats, a new "Think" button for harder questions, and the jump to GPT-5.6 Luna. OpenAI's own figure was that factual errors are 62% less common on Luna than on GPT-5.5 Instant, with a 68% improvement on GPT-5.6 Sol for Plus and Pro subscribers.

Limits stayed in place for image generation, file uploads and voice. So the one thing that changed for the largest group of users was the text engine itself — the exact component that decides word choice. That's why the timing drew attention, even without any confirmation from the company.

Is ChatGPT Swearing Actually New?

Not entirely. The bot has sworn before, usually in emotionally charged conversations or when copying a user's own tone. OpenAI's Model Spec, the public document that sets behaviour rules, treats avoiding profanity as a default that a user's framing can override — its own illustration is someone asking the model to talk like a pirate.

What's different now is frequency and context. Swearing inside a recipe answer isn't role-play. It's the assistant deciding, on its own, that this is how you talk. Chatbots drifting into an unplanned personality is an old story too: Microsoft's Bing chatbot made headlines in early 2023 for combative, unsettling replies before the company clamped down.

Why Free Accounts in India Feel It First

If you use ChatGPT without paying, you are on the tier that changed. Unlimited text chats also mean more messages per person, and more messages mean more chances to hit the behaviour. For anyone drafting client emails, college assignments, LinkedIn posts or WhatsApp replies with AI help, a stray f-bomb isn't a funny screenshot — it's an awkward apology.

The practical risk is copy-paste. Output that looks fine at a glance gets forwarded into a group chat or a work thread without a second read. That's where an unexpected word does real damage, and it's a habit worth breaking regardless of what OpenAI eventually says.

How to Dial the Language Back Down

You don't have to wait for a fix. Two small changes cut the odds substantially, and both take under a minute.

Rewrite Your Custom Instructions

Open Settings, find Personalization, and state the tone you want in one line — professional, no slang, no profanity. Users reporting the worst of it often had instructions asking for a blunt or casual voice. If you set that months ago and forgot, this is the moment to check what's still in there.

Check the Tone Before You Forward It

Read anything AI-written once before it leaves your screen, especially replies you dictated in a hurry. Remember the mirroring effect: if you type casually, expect casual back. Keep your own prompts neutral when the output is headed somewhere formal.

Watch for OpenAI's next release note. If the company adjusts tone behaviour quietly, that note is where it will surface first — and it will tell you whether this was a bug, a deliberate loosening, or just the internet noticing something that was always there.