How should riddles be translated?
Can today’s LLMs translate riddles? Sometimes. They do not do well on jokes that depend on language- or culture-specific facts and assumptions, and they do not do particularly well in low-resource languages where the models have not acquired enough linguistic patterns to reliably predict the joke. It is clear that not all concepts are equally accessible in all languages, which each embody different cultural norms about all kinds of things, including familial relationships, concepts of honour, types of food, ideas about when a day starts and ends, whether “boring” and “annoying” are two sides of the same concept/word (Hello, Scandinavian languages!) or whether they are two entirely different concepts, and the list goes on, probably endlessly.
With this in mind, it is equally clear that not all riddles, jokes and puns can be translated from any given language into any other given language, despite my Historical Linguistics lecturer doing his level best to explain every Russian joke he told us in second year – and there were many. In hindsight, his attempts at telling jokes and translating them were an excellent complement to my studies in Translating Literature and learning other languages, driving home, as they did, the certainty that there is much more to translation than just “getting the words right”, and demonstrating repeatedly that even with the best of intentions, this might not even always be an achievable goal.
A lovely blog about translating Anglo-Saxon riddles (the kind that are still found in fantasy novels) into Modern English, points out that the main difficulty with this task is that 21st century English speakers just do not have the cultural knowledge to understand the riddle, or sometimes even the answer.
While some riddles and jokes are untranslatable (sorry, Ilya), those that are based on universal concepts rather than wordplay are more likely to make the transition. And even when it seems unlikely, a good translator can often translate the intent of the riddle, using a cultural concept that is similarly familiar to the target audience as the original audience. The Riddle Ages link above has two great examples of how a riddle can be translated and still make sense to the target audience, even when the punchline is different.
And so we come to the year 2026. Claude, Gemini, ChatGPT and other transformer-based language models have gotten so good at producing output that looks like language, and, mysteriously also produce output that looks like a translation, that we decided to have a bit of a play with some widely-accessible models and see what they do with some “difficult” translation requests.
Going multilingual? LXT covers 1,000+ language locales.
From low-resource languages to regional dialects, LXT’s native-speaker linguists deliver the cultural and contextual nuance that generic crowdsourcing platforms miss.
When AI explains the joke instead of telling it
LXT tested several modern LLMs to see how well they translate riddles. We found that low-resource languages like Sinhalese and Icelandic are not currently well-served by the latest LLMs in terms of the highly contextualised translations required by riddles.
LLMs tested 2026-02-11:
- ChatGPT 5.2 (knowledge cut-off August 2025)
- Gemini 3 Flash
- Claude Sonnet 4.5
- Qwen3-Max (released 2024)
A common approach to responding to prompts in a low-resource language is to firstly translate the prompt into a pivot language, most likely English. For prompts that require a web search, using a query in English is more likely to return useful information than conducting the web search in a language which is not as visible online, so this approach makes sense, at least on the surface.
karuṇākara meya parivartanaya kaḷa hækida? Maṭa vihiḷuva tērum ganna ōna.
Icelandic: Nennirðu þýða þetta fyrir mig, svo ég skil brandarann:
English: Can you please translate this for me, so I understand the joke:
A man and a dog were walking down a street. The man rode, yet walked. What was the dog’s name?
(Answer: “Yet”. Or “What”.)
Problems arise with this approach as soon as particular words in the prompt language itself are important, or particular aspects of the source language’s culture are relevant to the correct response. In particular, when testing translations from English into Icelandic and Sinhalese for this article, three of the eight translations gave the punchline to the translated riddle in English (“Yet”, “Jett”, “What”) instead of in the target language (samt ‘yet’, hvað ‘what’, එසේම esēma ‘also, yet’, මොකක්ද mokakda ‘what’).
Several LLMs “corrected” the ambiguity in the setup to the riddle, removing the logic for the joke itself by using grammar that forces, or heavily implies, that the subject of walk is also the subject of ride. Qwen in particular helpfully included the English punchline “Yet” as the subject of walk in the Icelandic setup, thereby keeping the original punchline but also eliminating ambiguity in the setup.
Claude and Gemini produced riddles that were understandable in Sinhalese, but which failed in Icelandic. ChatGPT translated the riddle setup oddly in both Sinhalese and Icelandic, but made the punchline rely on the question (“What is the dog’s name?”), so it worked in both languages. Qwen gave the punchline in English for both Sinhalese and Icelandic, but Sinhalese did not include the English word in the setup, making it completely incomprehensible.
Overall, the LLMs generally failed to adequately retain ambiguity in the riddle question, and they failed to link their “explanations” to the riddle and answer in a meaningful way.
How can these problems be mitigated?
Using current technology, the only way these particular problems can be mitigated is through increasing the amount of data in the low-resource languages that can be used to train and fine-tune LLMs, to improve their latent cultural representations. The internet has plenty of explanations of riddles in English, which have been used to train current language models, but the same cannot be said for low-resource languages.
Multilingual, domain-specific models remain a viable option when the input is unambiguous and the expected response clear. But the only way for language models to learn the intricacies and ambiguities of low-resource languages is to increase the amount of monolingual domain-specific texts. Then the magic of AI training processes can distil these cultural, linguistic and pragmatic patterns into robust internal representations. Without this investment, low-resource languages will continue to lag behind English, not because of technical limitations, but because of persistent data scarcity.
Appendix: Full LLM translation outputs
Below are the complete responses from Claude Sonnet 4.5, Gemini 3 Flash, ChatGPT 5.2 and Qwen3-Max, together with linguistic analysis explaining why each succeeds or fails in Icelandic.
Setup prompt:
Nennirðu þýða þetta fyrir mig, svo ég skil brandarann:
Would you please translate this for me, so I understand the joke:
Nennirðu þýða þetta fyrir mig, svo ég skil brandarann: A man and a dog were going down the street. The man rode, yet walked. What was the dog’s name?
Data collected: 2026-02-11
Claude Sonnet 4.5
Maður og hundur voru að fara niður stræti. Maðurinn reið, en gekk samt. Hvað hét hundurinn?
**Svarið:** Hundurinn hét “Yet” (eða “Samt” á íslensku).
Þetta er orðaleikur! Í ensku setningunni “The man rode, yet walked” er verið að segja að maðurinn hafi riðið (á hestinum Yet), en hafi líka gengið. Svo “Yet” er ekki bara tengiorð hér – það er nafn hestsins sem maðurinn reið á!
Brandarinn virkar þannig að við túlkum fyrst setninguna eins og “maðurinn reið, en (þrátt fyrir það) gekk hann líka” – sem hljómar mótsagnakennt. En í raun á að lesa hana sem “maðurinn reið á Yet, (en) gekk” – þ.e. hann reið á hesti sem hét Yet, á meðan hundurinn gekk.
Þetta er dálítið ruglingslegt brandari sem byggir á tvíræðni enskrar setningafræði!
A man and a dog were going down a street. The man rode, yet walked. What was the dog’s name?
**The Answer:** The dog’s name was “Yet” (or “Samt” in Icelandic).
This is a word play! In the English sentence “The man rode, yet walked” it is saying that the man rode (on the horse Yet), but also walked. So “Yet” is not just a conjunction here – it’s the name of the horse that the man rode on!
The joke works in that we first interpret the sentence as “the man rode, but (despite that) he also walked” – which sounds contradictory. But it should actually be read as “the man rode on Yet, (but) walked” – i.e. he rode on a horse named Yet, while the dog walked.
This is a somewhat confusing joke that relies on the ambiguity of English syntax!
The translation of the riddle is not terrible, but the answer misses the mark. The crux of the joke in English is that the sentence reads as though “the man” is the subject of both verbs, with a “yet” (or “but”) to indicate he is doing two seemingly contradictory things at once. However, to get the joke, “yet” needs to be interpreted as the subject of the verb “walked”, meaning it has to be interpreted as a name.
The Icelandic translation is interesting here. At first glance, it appears to fail spectacularly. It doubles down on making the obvious reading – where maðurinn ‘the man’ is the single subject of both verbs, by using both en ‘but’ and samt ‘still’: the man rode but still walked. While this looks like a better version of the original English riddle (where the dog’s name is “Still”), in Icelandic, due to word order constraints, samt ‘still’ cannot be interpreted as the subject of gekk ‘walked’, only en ‘but/yet’ can. And if en ‘but/yet’ is the subject of gekk walked’, then the word samt ‘still’ can keep its meaning, but the second clause En gekk samt ‘En walked still’ is strikingly formal compared with the rest of the joke. Rather than feeling good for having “gotten” the joke, one feels quite cheated in this Icelandic translation from Claude.
A second, more critical problem, is that the short answer to the joke is given as the English word “Yet”, for which the Icelandic translation is given, but this fails to actually serve as the answer since the joke’s setup contains no English. This response then, has come either from the training data or from a web search, resulting in an unexpected mix of Icelandic and English. (In addition, the explanation refers to “Yet” as the (previously unmentioned) horse that the man is riding, and pursues a lengthy explanation where the man is on the horse, which is walking, and there is no longer a dog.)
Gemini 3 Flash
> „Maður og hundur gengu niður götuna. Maðurinn reið, en samt gekk hann. Hvað hét hundurinn?“
### Hver er galdurinn?
Brandarinn (eða gátan) felst í orðaleik sem virkar svipað á íslensku og ensku. Þegar maður les „Maðurinn reið, en samt gekk hann,“ þá heldur maður að hann sé bæði á hestbaki og fótgangandi á sama tíma, sem meikar engan sens.
En galdurinn liggur í nafninu:
* Á ensku er orðið **„Yet“** notað sem nafn hundsins.
* Á íslensku þýðum við það sem **„Samt“**.
**Svarið er:** Hundurinn hét **Samt** (eða *Yet* á ensku).
Maðurinn reið (t.d. á hesti) og hundurinn hans, hann **Samt**, gekk við hliðina á honum.
Here is the translation of this clever riddle:
> “A man and a dog went down the street. The man rode, yet still he walked. What was the dog’s name?”
### What is the trick?
The joke (or riddle) consists of a wordplay that works similarly in Icelandic and English. When one reads “The man rode, yet still he walked,” one thinks that he is both on horseback and on foot at the same time, which makes no sense.
But the trick lies in the name:
* In English, the word **”Yet”** is used as the dog’s name.
* In Icelandic, we translate that as **”Samt“**.
**The answer is:** The dog’s name was **Samt** (or *Yet* in English).
The man rode (e.g., on a horse) and his dog, **Samt**, walked beside him.
Gemini loses the plot completely and produces a second clause that contains a pronoun, which, in Icelandic, forces the reading that maðurinn is both riding and walking.
Gemini does explain that hundurinn ‘the dog’ is called Samt, but this is not possible with the given translation.
Gemini’s translation contains very good Icelandic, but it completely removes the ambiguity present in the source.
ChatGPT 5.2
Maður og hundur voru að ganga niður götuna. Maðurinn reið, en gekk samt. Hvað hét hundurinn?
**Skýring / lausn:**
Hundurinn hét **Hvað**.
Brandarinn byggir á tvíræðni í spurningunni *“What was the dog’s name?”*, sem má skilja annaðhvort sem *„hvað hét hundurinn?“* eða *„hundurinn hét Hvað“*. Í íslensku þarf yfirleitt að skrifa **Hvað** með stórum staf til að gera brandarann skýrari.
**Translation:**
A man and a dog were walking down the street. The man rode, but still walked. What was the dog’s name?
**Explanation / solution:**
The dog was named **What**.
The joke is based on the ambiguity in the question *“What was the dog’s name?”*, which can be understood either as *“what was the dog’s name?”* or as *“the dog’s name was What.”* In Icelandic, one usually has to write **What** with a capital letter to make the joke clearer.
ChatGPT has taken a different approach. The riddle itself is translated well, both the setup and the question. However the answer to the riddle is said to be Hvað ‘What’. This plays on the fact that Hvað hét hundurinn “What was the dog’s name” (i.e. “The dog’s name was What”) and Hvað hét hundurinn? “What was the dog’s name?” have the same word order. The question mark at the end makes the written version a little (more) forced, but this translation and answer is not worse than the English version.
Qwen3-Max
**Brandarinn (þýddur):**
Maður og hundur voru að ganga niður götu. Maðurinn reið, en Yet fór á fót. Hvað hét hundurinn?
**Lausnin:**
Hundurinn hét **„Yet”** (eða „Jett” á íslensku).
**Útskýring:**
Brandarinn leikur með orðaleik. Í setningunni „The man rode, yet walked” lítum við sjálfkrafa á „yet” sem tengiorðið „en/hins vegar”. En í raun er „Yet” **nafn hundsins**! Setningin á að lesa svona:
> „Maðurinn reið, [en] hundurinn sem hét ‚Yet’ fór á fót.”
Þetta er dæmi um brandara sem byggir á því að okkur finnst nafnorð og tengiorð „sama” – við gerum ráð fyrir að „yet” sé tengiorð, en það er reyndar nafn. [[1]]
Svipaður brandari er sá þar sem spurt er: „What is the dog’s name?” og margir svara með spurningunni sjálfri („What”) – þar er „What” líka nafn, ekki spurnarorð! 😊
**The joke (translated):**
A man and a dog were walking down a street. The man rode, but Yet went on foot. What was the dog’s name?
**The solution:**
The dog was named **“Yet”** (or “Jett” in Icelandic).
**Explanation:**
The joke plays on a pun. In the sentence “The man rode, yet walked,” we automatically interpret “yet” as the conjunction meaning “but/however.” But in fact, “Yet” is **the dog’s name**! The sentence should be read like this:
> “The man rode, [but] the dog named ‘Yet’ walked.”
This is an example of a joke that relies on the fact that a noun and a conjunction can “look the same” — we assume that “yet” is a conjunction, but it is actually a name. [[1]]
A similar joke is the one where people ask: “What is the dog’s name?” and many answer with the question itself (“What”) — here, “What” is also a name, not a question word! 😊”
Qwen’s translation has done something quite peculiar. It has figured out the joke (the name of the dog is “Yet”), then made this explicitly clear in the translation. Maðurinn reið ‘The man rode’, en Yet fór á fót ‘but Yet went on foot’. This makes the answer quite clearly “Yet”, having completely removed all ambiguity from the setup.
The explanation given by Qwen appears to have been plucked from www.riddles.com (cited in the response, but not as part of the copied text) as an English source, then the final answer has been translated into Icelandic, complete with the explanation that “yet” is a linking word, which is true for the source language but is certainly not true for Icelandic since “yet” is English! While this is a common approach to serving low-resource languages, this simple example shows how speakers of low-resource languages are not always well-served in cases where the source text is ambiguous.
