We Typed "idk tbh, hbu?" Into a Translator and It Came Back West Frisian
We Typed "idk tbh, hbu?" Into a Translator and It Came Back West Frisian
Here is the sentence: idk tbh, hbu?
Any English speaker under fifty reads that instantly. It means "I don't know to be honest, how about you?" Three abbreviations, one question mark, no ambiguity.
We put it through Translation Mixer, routed through French, and this came back:
In: idk tbh, hbu?
Out: Earlik sein, ik wit it net, of wol?
That is West Frisian, a language spoken by around half a million people in the northern Netherlands. Nobody asked for Frisian. Frisian was not in the chain.
Then we added one ordinary word to a similar sentence, and the problem vanished completely. What happens in between is the most interesting thing we have found in a while.
All results below are real output from Translation Mixer, collected in August 2026 using Google Translate as the engine.
The machine never knew you were writing English
Translation Mixer sends your text out through a chain of languages and then brings it home to whatever language you started in. To do that, it has to work out what language you started in. That detection step is where text-speak dies.
"idk tbh, hbu?" contains no English words. It contains three tokens that happen to be English abbreviations, and a language detector looking at raw letters sees short Germanic-looking strings with unusual consonant clusters. It guessed Frisian. Everything downstream was correct given that guess, which is why the sentence came home to Frisian instead of English.
We ran 49 abbreviations through on their own. Fifteen of them were not identified as English:
| You typed | Detected as | Came back as |
|---|---|---|
| icymi | Haitian Creole | krèm glase (ice cream) |
| tldr | Samoan | aotelega |
| lmk | Indonesian | beri tahu aku |
| istg | Maltese | Naħlef |
| ikr | Norwegian | Jeg vet |
| np | Polish | Na przykład |
| wyd | Afrikaans | 広い (wide) |
| omw | Dutch | rondom (around) |
| hbu | West Frisian | hbu |
| hmu | Mizo | en |
| yw | Welsh | Dwyrain (east) |
| ez | Basque | Ez |
| op | Dutch | Bovenop (on top) |
| rofl | Croatian | lol |
Mizo, for the record, is a Tibeto-Burman language of northeast India with about 800,000 speakers. It is the machine's best guess at "hmu."
Half of these are not mistakes
Look closer at that table and a pattern separates out. Several of those detections are simply correct.
wyd is a real Afrikaans word. It means wide. The machine read it as Afrikaans and translated it to Japanese as 広い, which also means wide. Nothing malfunctioned.
np is standard written Polish. "Na przykład" means "for example," and Poles abbreviate it "np." the way English writers use "e.g." Two languages independently invented the same two-letter shortcut for different phrases.
ez means no in Basque. op means on or up in Dutch.
So when a translator turns "np" into "for example," it is not failing to understand internet English. It is understanding Polish, correctly, because you gave it two letters and no reason to prefer one reading over the other. English text-speak has been colonizing a namespace that was already occupied.
It knew what you meant. It answered in Maltese.
The funniest category is the one where the machine got the meaning exactly right and delivered it in a language you do not speak.
istg stands for "I swear to God." Detected as Maltese, it came back as Naħlef, which is Maltese for "I swear." lmk returned as beri tahu aku, Indonesian for "let me know." tldr passed through Japanese as 要約, the correct word for summary, then went home to Samoan.
The knowledge is in there. The abbreviation was decoded properly. It was just decoded by a system that had already committed to the wrong language and had no way to reconsider.
Only one in the set got the meaning genuinely wrong, and it got it wrong in the best possible way. icymi was read as Haitian Creole and rendered into Japanese as アイスクリーム. Ice cream.
lol is the exception, and it explains the rule
One piece of text-speak survives everywhere, because it is the only one that has been separately reinvented in every language on earth.
French does not translate "lol." It swaps in mdr, short for mort de rire, dying of laughing. Japanese uses 笑, the character for laughter. Korean uses 웃음. These are not translations, they are local equivalents, and they exist because every internet culture needed a laugh token and built one.
That means the round trip depends entirely on whether the target language's laugh token maps cleanly back. We sent "lol" through sixteen languages, one hop each:
| Language | lol came back as |
|---|---|
| German, Portuguese, Korean, Dutch, Italian, Polish, French, Japanese | lol |
| Chinese | Ha ha |
| Thai | Haha |
| Hebrew | Hahaha |
| Spanish | ha ha ha |
| Vietnamese | laugh |
| Hindi, Indonesian | laugh out loud |
| Turkish | laughing out loud |
| Russian | I can't stop laughing |
| Arabic | Very funny |
That is a volume knob. The same three letters come back as a shrug in half of Europe, a full unpacked phrase in South Asia, and an unhinged confession in Russian. Arabic produced the only genuinely dangerous one: "Very funny" is what an English speaker types when something is not funny at all. A friendly laugh becomes sarcasm, with no warning to either party.
Context fixes almost everything
Now the good news, which is also the part that reframes the whole experiment.
Every misreading above happened to an abbreviation sitting alone. Put the same abbreviation in a sentence with real words and the detector locks onto English. Suddenly the results turn respectable:
| Alone | In a sentence |
|---|---|
| icymi → ice cream | "icymi, the meeting moved to Thursday." → "In case you missed it, the meeting has been postponed until Thursday." |
| wyd → wide | "wyd tonight?" → "What are you doing tonight?" |
| omw → around | "omw, be there in 5." → "I'm on my way, I'll be there in 5 minutes." |
| iirc → "I'm sure" | "iirc she said the deadline was Friday." → "If I remember correctly, she said the deadline was Friday." |
Every one of those is a correct expansion. The word "tonight" is doing an enormous amount of work in the second row: it is the only real English in the sentence, and it is enough.
Which is why "idk tbh, hbu?" ended up in Frisian. There was no anchor. A message made entirely of abbreviations gives the detector nothing to hold, and it is a rarer kind of message than you would think, because most people type at least one real word.
Note the iirc row too. On its own, iirc came back through Japanese as "I'm sure." In a sentence it stayed "If I remember correctly." A hedge became a certainty when isolated, which is the opposite of what the writer meant.
You get your text message back as an email
Meaning survives with context. Tone does not.
In: ngl, this restaurant slaps.
Out: Honestly, this restaurant is great.
In: brb, gotta walk the dog lol
Out: I'll be back, I have to walk the dog lol
Nothing there is wrong. "Gotta" became "have to," "slaps" became "great," and the four-character opener became a full adverb. Text-speak is compression, and translation is decompression, so a round trip is a one-way trip: you cannot get the short form back.
The register shift can change your meaning outright:
In: fwiw, I thought the first draft was better.
Out: For your information, I thought the first version was better.
"For what it's worth" is an apology for having an opinion. "For your information" is a correction. The same sentence went from deferential to slightly hostile, and the only thing that changed was an abbreviation being spelled out.
Five hops
Finally, the full mixer treatment. We sent a normal text message through French, Japanese, Korean, Arabic, and Hindi, and watched "lol" decay along the way: it held as lol in French, became 笑 in Japanese, 웃음 in Korean, then يضحك in Arabic and हंसते हुए in Hindi, both of which mean "laughing" as a description rather than a token.
By the end it was a stage direction in a screenplay, and the message came home in Mizo.
One more, because it ends on the right joke. We sent "ngl icymi the party got moved, wyd?" through Japanese, Korean, Chinese and Arabic:
Out: Honestly, the meeting place has changed, but what are you all doing?
All three abbreviations resolved correctly. The word that broke was "party," which became a meeting place. The slang made it home and the plain English did not.
Try It
The five-hop chain is loaded and ready:
→ Send your own text-speak through the mixer
Try it once with only abbreviations, then again with one real word added, and watch which language it decides you were speaking.
Related: Thunder Mug to Rizz runs the same experiment on spoken slang, old and new. We Round-Tripped One Paragraph Through 20 Languages has the damage data for the languages used in these chains. And Why Machine Translation Humor Works explains why ice cream is a funnier answer than a correct one.
Try it yourself →
Send your own sentence through the translation telephone game and see what comes back.
Related Posts
What Machine Translation Does to Slang, Old and New
Old slang was supposed to survive and new slang was supposed to break. Both held up about equally. What actually decides it is whether the slang is one word or a metaphor, plus what a single hop does to y'all and ain't.
What Language Is Hardest for Google Translate? We Tested 20
One paragraph, 20 languages, one round trip each, scored for damage. Japanese won, Finnish placed 19th, and re-running the whole test on a second engine changed the ranking, which turned out to be the most useful result.
Why Machine Translation Humor Works
Why do translations get funny? The real linguistics behind it: incongruity theory, compounding drift, and why the machine's mistakes are funnier because they aren't mistakes at all.