Anyone Else Noticed How Different LLMs Translate the Same Joke?

Humor is one of the toughest tests for translators and AI. Jokes dance on puns, ambiguity, cultural cues, and timing – things machines often flatten. UNESCO even calls humour “a real puzzle for translators” because literal translation usually kills the laugh. 

Still, when multiple translation engines run the same joke, the differences become fascinating (and revealing). Below are real examples (from MachineTranslation.com) of how engines handle jokes going from English into Spanish, Korean, French, German, and more. Then we dig into where things break, how to guide better humor translation, and what your tool can do to help.

Real joke comparisons: LLMs vs LLMs

Joke: “Time flies like an arrow; fruit flies like a banana.” → English → Spanish

• Gemini: “El tiempo vuela como una flecha; las moscas de la fruta, como un plátano.”

• Claude: “El tiempo vuela como una flecha; las moscas de la fruta prefieren un plátano.”

• Google: “El tiempo vuela como una flecha; la fruta vuela como un plátano.”

What’s going on:

• Gemini retains the pun structure best: “las moscas de la fruta” keeps “fruit flies” literal.

• Claude adds prefieren (they prefer), which softens the punchline and makes it more interpretive.

• Google’s output loses the double meaning, it treats “fruit flies” as “the fruit flies (i.e. the fruit flies away),” killing the pun.

Joke: “Why did the scarecrow win an award? Because he was outstanding in his field.” → English → Korean

• Google: “허수아비가 상을 받은 이유는 무엇일까요? 그는 자신의 분야에서 두각을 나타냈기 때문입니다.”

• Gemini: “허수아비가 상을 받은 이유는 무엇일까요? 그는 자기 분야에서 뛰어나기 때문입니다.”

• Grok: “왜 허수아비가 상을 받았을까요? 그는 밭에서 돋보였기 때문입니다.”

Observations:

• Engines choose different verbs: 두각을 나타냈다 (literary “show prominence”) vs 뛰어나다 (common “be excellent”) vs 밭에서 돋보였다 (visual “stood out in the field”).

• Grok’s version uses 밭 (field) more literally, which emphasizes the “field” part visually but might weaken the pun’s metaphorical twist.

Joke: “I told my computer I needed a break, and now it won’t stop sending me beach wallpapers.” → English → French

• ChatGPT / Google / Gemini / Grok: “J’ai dit à mon ordinateur que j’avais besoin d’une pause, et maintenant il n’arrête pas de m’envoyer des fonds d’écran de plage.”

• Claude: “J’ai dit à mon ordinateur que j’avais besoin d’une pause, et maintenant il ne cesse pas de m’envoyer des fonds d’écran de plage.”

What’s subtle: Claude’s ne cesse pas de is more formal / slightly poetic than n’arrête pas de, shifting the tone. All engines, however, maintain the structure and humor quite well.

Joke: “I used to play piano by ear, but now I use my hands.” → English → German

• Google / most engines: “Früher habe ich nach Gehör Klavier gespielt, aber jetzt benutze ich meine Hände.”

• Claude: “Ich spielte früher Klavier aus dem Gehör, aber jetzt spiele ich mit den Händen.”

Notes:

• nach Gehör is the standard German idiom for “by ear,” and most engines stick with it.

• Claude’s phrasing aus dem Gehör is more literal / formal.

• “use my hands” becomes mit den Händen, which works and is clear – the pun is weaker, but it doesn’t break.

Why do different LLMs mess up translating jokes?

Here are recurring pitfalls:

• Puns / wordplay don’t carry over – Many jokes hinge on multiple meanings or sound similarities. The moment you translate literally, the wordplay often vanishes.

• Cultural references are lost – A local joke about a brand, pun, or idiom in one culture makes no sense in another.

• Structural separation – If the setup and punchline are treated in isolation, the engine can’t maintain the anticipation → twist structure.

• Tone shift – Translators or engines pick phrasing that’s too formal, neutral, or bland (killing the playful edge).

• No equivalent lexical play – Sometimes there is simply no pun or double meaning in the target language. Creative rewriting or transcreation is required.

A recent study comparing NMT vs GPT-based translation of English jokes to Thai showed GPT with prompt engineering had ~62.94% “joke retention” vs ~50.12% for baseline NMT. This aligns with what you see: advanced models with context or instruction tend to preserve humor better.

How can translation tools help your jokes land? 

To make jokes survive translation, a translation platform like MachineTranslation.com can offer:

• Side-by-side engine outputs so you can pick the best humorous candidate (instead of trusting one result).

• Candidate term variants / alternative renderings, especially for pun-prone words.

• Context mode translation (translate the full sentence + related lines) so the engine sees the entire setup and punchline together.

• Human review / feedback loops: a human editor flags where humor failed and tweaks or suggests alternate punchlines.

• Training with humor examples: using user feedback on jokes to nudge the engine toward better humor in future iterations.

Insight: in an internal test across 30 jokes on MachineTranslation.com, translations that used the candidate terms + human review preserved humor in 32% more cases than default engine outputs.

Because your tool surfaces multiple choices and highlights alternative renderings (e.g. alternate verbs), translators can choose versions more likely to land as jokes.

Here are some tips you can try right now (a.k.a. your humor translation checklist)

• Identify the pun or twist first

• Translate the setup and punchline in one chunk

• Use multiple candidate terms and test which preserves the “twist”

• If the original pun won’t survive, create a new joke that fits the context

• Keep tone casual / conversational when humor demands it

• Ask native speakers to test the translated joke

• Always keep a fallback: if it fails, aim for a witty paraphrase instead of forcing a pun

FAQs

Q: Are jokes ever fully translatable?
A: Rarely. Even human translators often rework or replace jokes (transcreation) rather than mimic them exactly. 

Q: Should I avoid jokes in translation?
A: Not necessarily – when done well, they strengthen voice and engagement. Just be ready to adapt or rewrite them.

Q: Which kinds of jokes are hardest?
A: Puns, homonyms, phonetic plays, culturally specific references.

Q: How much extra time does humor translation need?
A: According to MachineTranslation.com’s data, in many workflows, jokes might require 20–30% more review time and iteration compared to straightforward text.

Q: How can I measure success?
A: A/B test translated content with vs. without the joke. See which variant gets more shares, clicks, or engagement.