My son sent an AI chatbot the message "oh great, MY CAR broke down AGAIN, exactly what I needed today" just to see what would happen, half expecting it to cheerfully respond with "That's wonderful news!" like some old customer service bot would have. It didn't. It picked up on the frustration immediately and responded with something sympathetic about car trouble timing being the worst.
That surprised him, and it surprises most people the first time they test it. Modern AI chatbots are actually pretty good at reading sarcasm, especially the more obvious kind — all caps, exaggerated enthusiasm about something clearly bad, that sort of thing. They pick up on it from the same context clues a person would: word choice, punctuation, the situation you're describing.
Where it gets shakier is the dry, deadpan kind of sarcasm that even people miss sometimes over text without a raised eyebrow or a tone of voice to go on. If you write something completely flat and sarcastic with no exaggeration at all, there's a real chance it takes you literally, the same way a friend reading your text without knowing your voice might. That's not a flaw unique to AI — it's the same problem sarcasm has always had in writing.
The practical takeaway: if you're being sarcastic with an AI and it matters that it gets the joke, either make it a little more obvious than you would with a close friend, or just say afterward, "that was sarcastic, by the way." It won't be offended, and it'll adjust. It's a strange kind of relationship to be coaching a computer program on your sense of humor, but here we are.
If you want the slow, no-jargon version of all this — the stuff I wish someone had handed me before I fumbled through it myself — that's exactly what I built On Ramp AI for. Check it out here.