Why is ChatGPT swearing more lately? Users say it is becoming less like a customer-service bot

The bottom line

ChatGPT, which is as polite as customer service, was recently discovered by users that it can suddenly curse.

Business Insider reported on August 18 that multiple users discussed on social platforms and Reddit: ChatGPT has been more prone to swearing recently, and even inserted f-bomb into recipes, GitHub fault explanations, and network protocol discussions.

The time point roughly corresponds to OpenAI switching the free user default model from GPT‑5.5 to GPT‑5.6 Luna. OpenAI did not confirm that this was a deliberate style change, nor did it explain why this round of swearing occurred.

So the most accurate title of this article is not "OpenAI secretly teaches ChatGPT to curse", but: Users observed that it became more human-like, but no one knew whether this was model style, personalization, or a brief fluctuation in behavior.

From "occasionally bursting out" to "why are you suddenly so skilled?"

ChatGPT has not been completely free from profanity in the past: It has occasionally used swear words when users complained, emphasized their tone, or turned on more casual personalization settings.

What makes users feel different this time is the frequency and scenarios. One user shared a recipe that included the f-bomb, and another answer about a glitch in Microsoft's code hosting service included "GitHub is having a shitfit." Some users said that the model began to imitate their own mantras and even generated expressions such as "big ass kolaches."

This changes from "can the model curse" to another more interesting question: Is it understanding your tone, or is it over-reflecting the way you speak?

The place that looks most like a real person may also be the most uncomfortable.

Some users like the change and feel that conversations feel more natural; one Australian user even said that more swear words make ChatGPT feel more like a real person.

But some people feel that it is excessive at first and then weakens. Personalization, user tone, model version, and context can all affect results.

This shows that "personalization" does not only have advantages. The more a model can imitate a user, the more likely it is to amplify the user's temper, prejudices, and expression habits; this is humor in chatting with friends, but it may become a risk in work emails, customer service replies, and children's usage scenarios.

It’s not OpenAI that turned ChatGPT into “rogue mode”

The current public evidence only includes user observations and media interviews, and OpenAI has not confirmed the existence of unified product changes. A few screenshots cannot be regarded as a statistical conclusion for all users, nor can differences caused by personalized settings, prompts, or context be ruled out.

But such small changes are worth recording, because they often reveal changes in model experience earlier than the parameters at the press conference: users are not testing benchmarks, but suddenly feel "it speaks differently today" while chatting.

Our judgment: AI’s “personality” is becoming a product indicator

Model manufacturers have been pursuing more natural, more interesting, and more human-like conversations. The problem is, being natural doesn’t mean being likeable, and being human doesn’t mean being reliable.

If you find that ChatGPT suddenly becomes too vulgar, you can check your personalization settings, change the system prompt to one that explicitly requires a professional tone, and manually review important output before doing so. Don’t regard a hot post on a social platform as a permanent revision of the model.

What to watch

• Whether OpenAI explains the relationship between GPT‑5.6 Luna and this round of behavioral changes.
• Whether foul language recurs reliably across accounts, personalization settings, and regions.
• Will the boundaries around models mimicking user tone become new security and brand controls?