Oh Wonderful, Another Thing AI Can't Actually Do Properly
What happens when you train a literal-minded algorithm to interpret human bitterness, and what I learned after wasting several hours testing it
I was sat in a coffee shop last Tuesday, sipping my cappuccino from a cup the size of a foot bath, and attempting to use voice dictation on my phone to send a withering text message to my mate. “Oh brilliant,” I muttered into the device, dripping with enough sarcasm to corrode metal, “another genius idea.”
The phone, bless its literal silicon heart, transcribed it perfectly and added a smiley face emoji. It was a perfect reversal of meaning, courtesy of technology that heard every word but understood absolutely nothing.
Which is exactly the problem I keep running into when I’m sarcastic with AI (sarcasm is in my DNA. Don’t judge me).
Teaching Machines to Detect Eye-Rolling
So can AI detect sarcasm? The short answer is: sometimes, sort of, maybe, if you’re lucky and the stars align. The longer answer is that AI’s relationship with sarcasm is like watching a puppy try to eat soup. It’s making an effort, there’s definitely some contact happening, but the fundamental mechanism is all wrong.
I spent an embarrassing amount of time last month testing this with Claude, ChatGPT, and a handful of other AI tools, because I’m the sort of person who has nothing better to do than torment algorithms with increasingly bitter statements to see if they flinch. “Oh wonderful, another email,” I’d type. “Fantastic, exactly what I needed, more admin.” “Brilliant, cheers for that.”
The results were about as consistent as British weather. Sometimes the AI would catch it, acknowledge the sarcasm, and respond appropriately. Other times it would take me at face value and offer helpful suggestions for managing my apparently delightful inbox. It was like talking to someone who doesn’t get the joke when you’re clearly taking the piss.
How Machines Parse Your Bitter Disappointment
Here’s the technical bit, stripped of the marketing waffle: large language models don’t “understand” sarcasm the way humans do. They recognise patterns. They’ve been trained on billions of text examples where certain linguistic markers, combined with specific contexts, tend to indicate sarcasm. Words like “wonderful” paired with obviously negative situations. Exaggerated politeness in response to mild annoyances. The textual equivalent of someone sighing heavily while saying “fine.”
But after wasting several perfectly good hours on this, teh AI isn’t detecting your tone. It’s making a statistical guess based on pattern matching. It’s seen “Oh great, another meeting” tagged as sarcastic in its training data enough times that when you type something similar, it thinks “probably sarcasm.” But change the phrasing slightly, remove the contextual clues, or use a less common sarcastic structure, and the whole thing falls apart like a cheap umbrella in actual weather.
The models get better with more context. If you’ve been having a conversation where you’re clearly annoyed about something, and then you say “Oh brilliant, more of that,” the AI can piece together that you’re not actually thrilled. But drop that same phrase into a fresh conversation with no context, and you’re rolling the dice on whether it interprets you correctly or offers you congratulations.
I discovered this the hard way when I asked an AI to help me draft responses to some particularly tedious correspondence. I’d say something like “Oh lovely, another passive-aggressive note from the neighbours,” expecting it to help me craft an equally frosty reply. Instead, it generated a warm, friendly response thanking them for their consideration. Absolutely useless. The AI had heard the words but completely missed the subtext, like a tourist nodding along to banter they don’t quite follow.
The Bit Where I Actually Help You
Right, here’s what I actually do now when working with AI, having learned that expecting it to automatically detect sarcasm is like expecting your microwave to appreciate irony.
Be Explicit About Your Tone When It Matters
I know, I know. Having to explain that you’re being sarcastic defeats the entire bloody point of sarcasm. But if you need the AI to understand your actual meaning, you’re better off just telling it. “I’m being sarcastic here” or “This is bitter irony, not genuine enthusiasm” works far better than hoping the algorithm picks up on your textual eye-roll.
It feels ridiculous, like having to explain a joke. But the alternative is watching the AI take your caustic observation at face value and respond with genuine cheerfulness, which is somehow even more depressing.
Use Sarcasm Markers the AI Actually Recognises
After extensive testing (again, I have too much time on my hands), I’ve found that certain structures work better than others. The classic “Oh great” or “Oh wonderful” combined with obviously negative content. The exaggerated “Just what I needed” before describing something annoying. “How delightful” followed by something objectively not delightful.
These are the linguistic equivalent of putting your sarcasm in high-visibility clothing. The AI has seen these patterns enough that it’s more likely to catch them. Is it authentic human communication? No. Does it work? More often than subtlety does.
Give Context Before Dropping Sarcastic Bombs
If you’re working on something where the AI needs to understand your actual sentiment, build up the context first. Tell it you’re frustrated, annoyed, or deeply unimpressed before launching into sarcastic commentary. “I’m incredibly irritated by this situation” followed by “Oh brilliant, more complications” gives the AI enough clues to piece together that you’re not actually celebrating.
It’s like the difference between texting “great” to someone who knows you well versus texting “great” to a stranger. One understands you mean the opposite, the other thinks you’re pleased. The AI is always the stranger until you give it context.
Accept That Sometimes It’ll Just Miss It Completely
There’s a certain type of bone-dry, understated sarcasm that AI will never catch. The sort where you say something that could technically be positive but is clearly not, delivered with the textual equivalent of a flat stare. “That’s interesting” when something is catastrophically boring. “I’m sure that’ll work out fine” when disaster is obviously imminent.
If you use that style of sarcasm (and honestly, it’s the best kind), just accept that you’ll need to rephrase or explain when working with AI. Or don’t, and enjoy the occasional surreal moment when the AI responds with genuine optimism to your profound pessimism. There’s a certain grim humour in it.
The Disappointing Circle Back
In the coffee shop, I eventually gave up on voice dictation and typed my message manually, adding enough profanity that even the most literal algorithm would have got the gist. The AI had perfectly transcribed my words but completely butchered my meaning, which is basically the same problem humans have been having with technology since we invented the automatic checkout that shouts at you for unexpected items in the bagging area.
Can AI understand sarcasm? Sometimes. Sort of. When you make it embarrassingly obvious. Which rather defeats the point, but then again, so does most of what we do with technology.
At least now you know how to work around its limitations, even if explaining your bitter irony to a machine feels about as satisfying as explaining a punchline to someone who’s already nodding along, pretending they got it five minutes ago.
Ask The Hum
Question: My dad has started using voice assistants for everything and gets furious when they don’t understand him. He spent twenty minutes shouting “ALEXA, PLAY QUEEN” before giving up and blaming “the bloody thing.” He won’t accept that maybe he needs to enunciate. How do I help without being condescending?
Jamie, Glasgow
Dear Accent Mediator,
You can’t help without being condescending because the entire situation is condescending from every possible angle. Your dad’s absolutely right that the bloody thing doesn’t understand him, because Amazon trained Alexa on approximately four million hours of Californian podcast hosts saying “awesome” and roughly twelve minutes of actual Scottish people existing.
The voice recognition wasn’t designed for him, it was designed for a fictional American who pronounces every vowel like they’re in a corporate training video. But he’s also completely wrong about the solution, which is that he now lives in a world where a plastic cylinder gets to decide whether his accent is legitimate enough to deserve music.
The technology is rubbish, but so is his refusal to accept that he’s been handed a system that requires him to perform a flattened version of himself to make a speaker work.
The real issue isn’t enunciation, it’s dignity. He’s spent seventy-odd years speaking perfectly clearly to actual humans, and now some American tech company has decided his voice doesn’t parse correctly, so he’s expected to either shout louder or cosplay as someone from Surrey. You’re watching him realise that convenience has been redefined as “works brilliantly if you sound like the target demographic, otherwise get fucked.”
He’s angry at the wrong thing, but at least he’s angry.
The Hum