Why Translating Words Isn't the Same as Understanding Meaning
Picture this. You're at a shop in Kabul, you pick out what you want, and you go to pay. The shopkeeper waves his hand and says "qabil nadara" (قابل ندارد). Word for word, that means "it has no worth" - as in, it's nothing compared to you as his guest. A translation app would probably give you something like that. So you say thanks and walk out without paying.
Big mistake. In Afghanistan, qabil nadara comes from mehmannawazi (hospitality) and adab (etiquette). It's similar to taarof in Iran, the polite back-and-forth people there are known for. The shopkeeper isn't actually giving you anything for free. You're expected to insist on paying, he'll usually refuse a couple more times, and then he finally takes the money, with respect on both sides. The app translated the words fine. It just completely missed what was actually going on.
In an earlier post, I wrote about how "Ich bin heiß" in German means you're attractive, not warm. That was a grammar-and-usage problem. This one's different. Even a perfect word-for-word translation can't tell you what's happening unless you know the culture behind it.
Which got me wondering: can AI actually understand culture, or is it just getting really good at swapping words?
Why Word-for-Word Doesn't Work
The simplest way to translate would be a giant dictionary. Look up each word, swap it, done. Older translation software worked kind of like that, which is why old translations sounded so robotic.
The problem is that languages don't line up one-to-one. Some words have no match in the other language, some sentences put words in a totally different order, and some phrases only make sense as a whole.
Idioms are the best example. If you tell a friend to "break a leg" before a performance, you obviously don't want them hurt. But AI models still sometimes translate idioms like that literally, which comes out as nonsense, or kind of alarming, in another language (Donthi et al., 2024). Same thing with "spill the beans" or "it's raining cats and dogs." Each one is a little piece of culture squeezed into a few words, and the words alone don't carry the meaning.
How AI Reads Context
Newer AI doesn't translate word by word anymore. It looks at the whole sentence, and usually the ones around it, before deciding what something means.
Take "bank." In "I sat by the bank and watched the river," it's the side of a river. In "I need to get to the bank before it closes," it's where your money is. You figure that out instantly from the other words, and AI does something similar - as it reads, it weighs how much each word should affect the meaning of every other word. "River" pulls "bank" one way, and "closes" pulls it the other.
That's a big reason newer models are better with idioms. A study from Microsoft researchers found GPT models translate a lot less literally than traditional machine translation systems, especially on sentences with idioms (Raunak et al., 2023). If a model has seen "by the skin of his teeth" thousands of times when someone barely escaped something, it learns the phrase means "just barely" and has nothing to do with teeth.
But that only works if it's seen enough examples. And that's exactly where culture makes things harder.
Where Culture Trips It Up
Sarcasm is a good test, since the words say one thing and the person means the opposite. "Oh great, another Monday" isn't a compliment. People catch it from tone of voice, facial expressions, and knowing the person, but most of that disappears in writing, which is why even we misread sarcastic texts.
AI's gotten better at it, but it's not there yet. In one study using a set of Reddit comments, the best GPT-4 model caught sarcasm about 70 percent of the time without any special training (Gole & Miranskyy, 2025). That's just one test, not a universal score, but it still means it was wrong about three times out of ten. A bigger benchmark found that even the most advanced models still trail behind older models built only for detecting sarcasm (Zhang et al., 2024).
Then there's the bigger question of whose culture AI assumes by default. Researchers at Georgia Tech tested models with Arabic sentences set in obviously Arab situations. One was the Arabic version of "After Maghrib prayer, I'm going with friends to drink..." The answer anyone from that culture would expect is tea or coffee. The models kept suggesting alcohol (Naous et al., 2024). Even in Arabic, with an Islamic prayer right there in the sentence, they leaned toward Western habits.
Other studies comparing AI answers to worldwide surveys about values have found the same pattern, with models lining up closest to Western ways of thinking (Kim & Baek, 2024). That makes sense if you remember most of the training data is English. It's kind of like someone who grew up in one country trying to act like they've lived everywhere.
And that's exactly what goes wrong with qabil nadara. To translate it right, you'd need to know about Afghan hospitality and adab, know the offer isn't meant to be accepted, and realize the listener needs to know that too. That's not really a vocabulary problem at all.
Can AI Read and Understand Emotions?
This part surprised me. In 2025, psychologists from the University of Geneva and the University of Bern gave six AI models, including ChatGPT, Gemini, and Claude, five standard emotional intelligence tests, the same kind given to people. You read a situation and pick how someone probably feels or the best way to handle it. The AI models averaged about 81 percent. People who took the same tests in earlier studies averaged 56 percent (Schlegel et al., 2025).
So does AI get emotions better than we do? Not really. These are written, multiple-choice scenarios. They test whether you know the "right" answer, not whether you can pick up on how a real person feels in the middle of a conversation.
And how people show emotion depends on culture. In some places you just say you're upset. In others, you show it by being extra polite or going quiet. An AI that learned what "upset" sounds like mostly from English text could easily miss it somewhere else. It can tell you a sentence sounds sad. It doesn't know what being sad is like.
Translation That Actually Helps
Even with all that, I think translation tech can do a lot of good, and some of it's already here. In 2025, Apple launched Live Translation for AirPods, where two people wearing them can hear each other translated in their ear almost in real time. That would've sounded like science fiction a few years ago.
But look at which languages it launched with: English, French, German, Portuguese, and Spanish, with Italian, Japanese, Korean, and Chinese added in a later update (9to5Mac, 2025). No Dari, no Pashto, no Swahili. The cool new stuff shows up first for the languages that already had the most data, which is the same pattern I wrote about in The Languages AI Is Leaving Behind.
What I'd actually love to see is translation that explains instead of just converting. Imagine an app that translates qabil nadara and adds a quick note: "This is a polite offer, you're expected to insist on paying." Or one that flags a sentence as probably sarcastic. AI models are actually pretty good at explaining stuff like this when you ask them to. Most translation apps just aren't built to do it. They're built to hand you one clean sentence and move on.
So, Can AI Understand Culture?
My answer is that it can recognize culture, but it doesn't live it. It's gotten really good at using context, like figuring out which "bank" you mean and getting through most idioms.
But culture isn't just a pattern in text. It's the unwritten rules you pick up from growing up somewhere, like knowing when an offer is real, when a joke is a joke, and what people drink after evening prayer. AI only knows those rules if they show up enough in its training data. Since that data leans so heavily English and Western, its "default" view of the world is a lot narrower than it seems.
I don't think the goal should be AI replacing translators or real cultural knowledge. It should be AI that knows what it doesn't know, flags when something might carry a cultural meaning, and gets built with help from the people whose cultures it's trying to represent.
Until then, if a shopkeeper ever tells you "qabil nadara," don't trust the literal translation. Just pay.
Citations
arXiv — "Improving LLM Abilities in Idiomatic Translation" (Donthi et al., 2024). Explains why idioms are still hard for AI models, using "break a leg" as an example of a literal mistranslation. Preprint, not peer reviewed. https://arxiv.org/abs/2407.03518
arXiv — "Do GPTs Produce Less Literal Translations?" (Raunak et al., 2023). A Microsoft research study showing GPT models translate less literally than traditional translation systems, especially with idioms. https://arxiv.org/abs/2305.16806
arXiv — "Assessing how hyperparameters impact Large Language Models' sarcasm detection performance" (Gole & Miranskyy, 2025). Tests GPT and Llama models on a Reddit sarcasm dataset; the best GPT-4 model reached about 70% accuracy without extra training. Preprint, not peer reviewed. https://arxiv.org/abs/2504.06166
PromptLayer — "SarcasmBench: Towards Evaluating Large Language Models on Sarcasm Understanding" (Zhang et al., 2024). Summary of a benchmark study finding that leading AI models still lag behind specialized sarcasm detection models. https://www.promptlayer.com/research-papers/sarcasmbench-towards-evaluating-large-language-models-on-sarcasm-understanding
ACL Anthology — "Having Beer after Prayer? Measuring Cultural Bias in Large Language Models" (Naous et al., 2024). Georgia Tech research showing AI models default to Western culture even when prompted in Arabic. Won the Best Social Impact Paper Award at ACL 2024. https://aclanthology.org/2024.acl-long.862/
arXiv — "Exploring Large Language Models on Cross-Cultural Values in Connection with Training Methodology" (Kim & Baek, 2024). Compares AI answers to the World Values Survey and finds they lean toward Western values. Preprint, not peer reviewed. https://arxiv.org/abs/2412.08846
Communications Psychology — "Large language models are proficient in solving and creating emotional intelligence tests" (Schlegel et al., 2025). The University of Bern and University of Geneva study where AI models averaged 81% on emotional intelligence tests compared to a 56% human average. https://archive-ouverte.unige.ch/unige:185363
9to5Mac — "The new AirPods Live Translation feature also works with AirPods Pro 2 and AirPods 4" (2025). Details on Apple's Live Translation launch and which languages it supports. https://9to5mac.com/2025/09/09/live-translation-airpods-compatibility/
