I tested three chatbots to see if AI can understand Carioca sarcasm

Last night, after finishing the month-end closing, I was sitting in front of my laptop. One in the morning, a breeze coming in from Copacabana, that hour when your head is no longer full of numbers but of pointless curiosity. I was tossing random things at a chatbot, as I do, when a question hit me: do these things actually understand the way I talk?

I speak Brazilian Portuguese, specifically the Rio variety, Carioca Portuguese. It's not just a dialect. It's a language where context and tone do half the work. Whether "Tá bom" actually means "okay" or "enough, shut up" depends entirely on the intonation and the situation. Over text, it gets even more ambiguous. So I grabbed ChatGPT, Gemini, and Claude and ran experiments over several days.

Testing Method

Nothing fancy. I picked sentences I actually use in everyday life: messages to coworkers, complaints to friends, short comments I might post on social media. The criteria were three things: can it catch a sarcastic tone, does it understand Carioca gíria (slang), and does it respond appropriately to the context?

For example, I entered a sentence like this: "Adorei ficar até meia-noite fechando balanço, melhor programa de sexta." It means I stayed until midnight on a Friday doing month-end closing, obviously not something I enjoyed. More of an accountant's self-defense mechanism.

The Results Were All Over the Place

ChatGPT was the best at catching sarcasm. It responded to the sentence above with something like "sounds like a lousy Friday," and kept the conversation going while maintaining the tone. It was a bit weak on distinctly Carioca slang, though. When I threw in "Vacilão," it gave me the standard Portuguese dictionary definition and missed the nuance the word carries in Rio (roughly "pathetic loser").

Gemini was the exact opposite. It knew the slang fairly well, but it often took sarcastic remarks at face value. When I wrote "Nossa, que maravilha trabalhar no sábado" (wow, working on Saturday is the best), it actually started listing the advantages of working on Saturdays. I was genuinely angry for a moment.

Claude landed somewhere in the middle. It detected sarcasm most of the time and tried to read the context, responding with things like "sounds like you had a tough time." But occasionally it leaned too hard into empathy, shifting into a therapist tone. What I needed wasn't consolation. I needed someone to be sarcastic right along with me.

All three models were weak when it came to long sentences mixing double negatives and irony. With a sentence like "Não é que eu não goste de pagar imposto, né" (it's not that I don't like paying taxes, right), all three either reacted a beat too late or interpreted it literally. Any Brazilian would pick up in half a second that this is a complaint about taxes.

So Who Won?

None of the three reached the core of Carioca sarcasm. ChatGPT was the best at detecting tone, Gemini came out ahead on vocabulary, and Claude had a slight edge on reading context, but its tendency to over-empathize held it back.

One thing that did impress me: when I told ChatGPT to "answer like a Carioca," the tone shifted noticeably. Its sarcasm detection rate went up, and it even started throwing in interjections like "Pô, foda." Sure, it would sound more natural coming from someone actually from Rio, but it was on the right track.

Something I kept thinking about throughout this experiment: an AI "knowing" a language and "knowing" a culture are two entirely different things. A model with flawless Portuguese grammar still can't get a joke from a back alley in Rio, because language is more than just a combination of words. It's similar to how numbers in an accounting ledger are meaningless if you don't understand the cash flow behind them.

It's still the early hours. My coffee has gone cold, and I have work again tomorrow. The chatbots understand about half of my sarcasm. The other half will probably remain a job for humans, at least for a while.