Thunder Mug to Rizz: We Ran a Century of Slang Through a Translator

We went in expecting a clean result. Old slang has been sitting in books and newspapers for a hundred years, so a translation model has seen it. New slang was invented on the internet three years ago, so it hasn't. Old should survive, new should fall apart.

That is not what happened. Both eras survived about equally well, and the failures in both came from the same place, which has nothing to do with when a word was coined.

Every result below is real output from Translation Mixer. Classic and modern terms went through the same chain, English to Japanese to Swahili to Finnish to Arabic and back, so the comparison is fair.

The Classics

Twelve terms, mostly for everyday objects, from roughly 1900 to 1970:

That is a better performance than we predicted. The machine knows that gams are legs and vittles are food, and it correctly read "the bee's knees" as praise rather than a body part. Note that it also quietly modernized the props: a record became an album, an icebox became a refrigerator. The meaning survives, the period does not.

Two of the twelve broke, and they broke in the same way:

In: Grandpa kept a thunder mug under the bed.
Out: My grandfather kept a thunderclap mug under his bed.

In: Somebody has to empty the honey bucket.
Out: Someone needs to pour water into a bucket of honey.

Both are chamber pots. The machine read both as compound nouns about weather and about honey. The second one is worse than a vocabulary miss, because the action reversed: emptying a chamber pot became pouring water into a container of honey. A chore became a recipe.

The third failure is the one worth framing:

In: You're getting too big for your britches.
Out: You're starting to get really smart.

An accusation of arrogance came back as a compliment.

The Modern Terms

Same chain, twelve terms coined mostly in the last decade:

The machine knows current slang better than most people over forty. It decoded sus, ghosted, flex, slaps, and cooked without trouble. "Bet" as agreement came back as "Absolutely," which is right, though the meeting time slid from seven to one o'clock somewhere in the chain for reasons we cannot account for.

The failures again cluster around phrases rather than words:

In: She ate and left no crumbs.
Out: He ate it all without leaving a single piece.

Read as a description of a meal. Similarly, "That outfit? It's giving main character" came back as "Those clothes? They perfectly showcase the main character's personality," which is a reasonable paraphrase of something nobody said.

And then the best result of the batch:

In: That guy has unspoken rizz.
Out: The man does not speak loudly, but he possesses a vast fortune.

Rizz, shortened from charisma, was read as riches. The Japanese hop rendered it as 相当な財力 (considerable financial power) and every language after that faithfully carried the money forward. The man is now wealthy and taciturn.

The Pattern Is Structure, Not Age

Line the failures up and the dividing line is obvious. It isn't old versus new.

Slang that is a single word standing in for another word survives. Jalopy, vittles, gams, icebox, sus, flex, slaps, cooked, ghosted. These are lexical substitutions, one item swapped for another, and that is exactly the operation a translation model performs. They have dictionary entries and the dictionary is enough.

Slang that is a metaphor assembled from ordinary words fails. Thunder mug, honey bucket, too big for your britches, ate, it's giving. Every word in those phrases is common and the meaning lives in the combination, not in any part. The machine translates the parts correctly and the meaning falls out.

This is the same failure we documented in loanwords and in the telephone game archive. What breaks is never the unfamiliar word. It is the familiar words doing unfamiliar work.

How "Mid" Became "Very Good"

One modern term inverted completely, and because we logged every hop we can show exactly where:

In: Honestly the new album is mid.
Japanese: 正直言って、ニューアルバムはまあまあだ。 (so-so)
Swahili: Kwa kweli, albamu mpya ni sawa tu. (just okay)
Finnish: Itse asiassa uusi albumi on ihan ok. (quite ok)
Arabic: في الحقيقة، الألبوم الجديد جيد جداً. (very good)
Out: Actually, the new album is very good.

No single hop is wrong. Japanese correctly read "mid" as mediocre. "So-so" to "just okay" is fair. "Just okay" to "quite ok" is fair. But "quite ok" to "very good" crosses the line from lukewarm to positive, and once it crossed, nothing pulled it back. Five small upward nudges turned a dismissal into a recommendation.

That is compounding drift with a direction, and it is worth knowing about if you ever machine-translate a review.

What Happens to Dialect

The most interesting results were not slang at all. We took nonstandard English and sent it through a single hop of Spanish, which is the gentlest test in this entire post.

Every sentence came back speaking standard English:

The double negative was resolved, "ain't" became "have not," "finna" became "going to," and Pittsburgh's "yinz" became a plain "you." One hop, and the regional voice is gone. Nobody instructed the machine to standardize anything. Standard English is simply what sits on the other side of the round trip, because that is what the training data mostly contains.

One sentence lost more than its accent:

In: Y'all fixin' to head out?
Spanish: ¿Ya se van?
Out: Are they leaving already?

Spanish "se van" can mean "you all are leaving" or "they are leaving," and the machine picked third person coming back. A question asked directly to a group became a question about other people who aren't there. The grammar is fine and the audience changed.

There is one survivor, and it is instructive:

In: It ain't over till it's over.
Out: It ain't over till it's over.

Not one word moved, "ain't" included. This is a famous Yogi Berra line, and the machine has seen it thousands of times as a fixed unit rather than as a sentence to parse. Quoted dialect in a famous sentence is safe. The same word in your own sentence is not.

Try It

The "mid" chain is loaded, since watching a word climb from mediocre to good over five hops is more fun than reading about it:

→ Run your own slang through the same chain

Regional terms work well here. So does anything your parents would need explained.

---

Related: We Typed "idk tbh, hbu?" Into a Translator does the same to abbreviations, where the failure has a completely different cause. Loanwords: 30 Words English Stole From Other Languages runs the same experiment on borrowed vocabulary. We Round-Tripped One Paragraph Through 20 Languages has the damage data for each language used in these chains. And Why Machine Translation Humor Works explains why "a vast fortune" is funnier than a correct translation would have been.