🅭🅯🄎

---

The Original

#

The Background

#

In my daily life, I often have to operate in three different languages: English, some weird German dialect very few people speak, and French. While I’m pretty much equally comfortable in all of these languages, I do have certain preferences. Basically, situations where I feel slightly more at ease in a particular language.

Writing, for instance, is something I prefer doing in English because, at this stage, all my reading almost exclusively happens in it. That includes watching movies in languages I don’t understand: the subtitles are always in English.

The Observation

#

I’ve been told by a few people I personally know that they’d love to read (some of) my work. Unfortunately, their English is almost non-existent. Well, they say unfortunately, but if you’re already familiar with any of my writings, I think we can both agree that the prefix un- can be dropped.

So in light of the above and somewhat unsurprisingly, I’ve been thinking about translating my own work for a while already. Making my writings accessible for people I have to interact with in person while updating their perception of me being slacker. After all, when they see me going for a walk in the middle of a workday while they’re busting their ass, something doesn’t add up. But, to be honest, with the kind of writing you can expect from me, translating my work into their language is probably not going to help when it comes to accessibility.

That being said, the reason I wanted to publish this piece (actually more so than I wanted to write it) is because translations will be a thing whether I take care of them myself or not. In other words, I have noticed massive improvements in translation capabilities in my own experiments with (local) large language models (LLMs) — at the time of writing (still) a bunch of static, digital brains that live on my own computer, in private and without a subscription or a requirement to share any of my data with anyone. The question: where is this going to lead over time?

The Evolution

#

Back when I was still earning a living as a professional translator, computer-assisted translation (CAT) was already quite useful. Over time, however, it became more and more obvious that the human factor in mediating between multiple languages will diminish until it just ceases to exist.

Indeed, the local LLMs I currently use are pretty much on par with a human translator. Mistakes and all. Proofreading just feels like it used to back at work. Reading translations from very competent translators — the people I was fortunate enough to work with.

So as curiosity got the better of me, I started to translate this website by outsourcing the grunt work to my self-hosted, static, digital brains. Ironic, I know. Describing translation as grunt work — coming from a former professional translator. That said, it doesn’t take a significant cognitive effort to see where all of this is going.

Since automatic translations are slowly becoming a part of our life, from today’s perspective it seems all but logical to wait a little longer and I won’t even have to bother using my local LLMs to translate this website. Essentially, it will be seamlessly displayed in the visitors’ preferred language — when the software enabling someone to visit this website (e.g. web browsers) automatically translates my work to whatever language the visitors feel comfortable in, so to speak.

Now, I’m actually cool with that. I mean, seamlessly removing language barriers? Sounds great! At the same time, there’s a subtle nuance — something that is going to be trapped inside the original. And it doesn’t matter whether my work is translated by a human or an AI. Something I cannot emphasize enough: from my perspective, this isn’t some human vs. AI kinda situation. In fact, it doesn’t even have anything to do with translation itself, but with how I personally approach each language in isolation (not counting the inevitable cross-pollination responsible for things like transfers and code-switching).

The Conundrum

#

Depending on the language I use, I do have a bit of a different style expressing myself. Not necessarily in terms of metaphorically engineering a thought process, mind you. My languages are far too close for that¹. Just tone and style.

Of course, instructing a translator — human or AI — to keep tone and style in the target language wherever feasible is no big deal. Works like a charm. However, when I instructed my local AI to translate Coexisting with Artificial Intelligence, I noticed that I would’ve never used the same tone and style if I had written the original in French or German. Especially in French because my ability to write in that language has dramatically declined over the years. I simply stopped writing in it and there isn’t really much of a reason to do so other than communicating with friends.

Anyway, the translations were great. Mostly correct as well as preserving tone and style of the English original. But from my point of view, something was off. Essentially, there was still a desire to rework them into something more “original” sounding — like I would express myself in the target language if it weren’t a translation.

To put this into perspective: depending on the piece of writing and its complexity, genre etc., I would focus on a balance between translation and adaptation, occasionally getting clarification from the original author. That relationship between the author, myself as the translator, and the author’s work, is somewhat different when the author and the translator are one and the same. For instance, the question regarding translation and adaptation can be broadened to include rewriting the piece from scratch in another language while retaining the core ideas and narrative. And most importantly, no need for a comparison with an original.

This is something I have realized especially in the case of Coexisting with Artificial Intelligence. The reason: that particular text is somewhat messy with overlong paragraphs, occasionally complicated sentences and an inconsistent tone and style². In short: it’s obviously flawed. An experiment of some sorts whose purpose was to linguistically draft impressions without worrying about neatly packaging them into easily digestible chunks. And that isn’t how I would’ve approached the piece in another language. To me, at least, it would feel somewhat unnatural to express myself in French or German in that particular style (regardless of what you might think as a reader). Instead, I would find other ways to “draft impressions without worrying about neatly packaging them into easily digestible chunks”. More suitable for each language.

Now, when it comes to my past publications, I’m just a reader as well. And it’s easy to see that experimenting as a writer doesn’t automatically translate to experimenting as a reader. That said, while I might sound harsh with some of my own writings, I would never mess with a text like Coexisting. At least not in its original form. Except fixing mistakes of any kind (I do care as much). After all, I can do whatever I want on my website, including rewriting and adapting imperfections. When they are deliberate, however, there’s no reason to do so. And that leads me back to translations.

The same imperfections — or occasional subversive writing — don’t seem to work in the other languages. At least not from my perspective. It’s almost as if I had to come up with new flaws that fit within my particular style in each language. Difficult not to think that in doing so I would actually start to mess with the original. Something I intend to avoid whenever possible.

It's a bit of a conundrum, I suppose.

Does that make these writings “untranslatable” in my eyes? Not really. My own style is just an emergent property from an activity that is far more important to me: shaping my thoughts into tangible objects. Besides, once published, they don’t belong to me anymore. Translated or not.

Indeed, this is somewhat related to the question of authorship and control. Personally, I was never too worried about control. For one, I’m fundamentally opposed to intellectual property — something I don’t want to waste time discussing as I can simply make my case by preserving the freedom for everyone to use my material as they see fit. By the way, that freedom does not prevent people from making financial contributions to projects, does it? As a response to those who, somewhat flabbergasted, may want to object: “But no money, no honey!”.

Second, my work is too “niche” and my writings are culturally too insignificant for me to be concerned with countless translations in different languages. In other words: I’m not writing bestseller material and, as such, there was never a reason for me to think about my work being translated or adapted into another language (unless I take care of my own translations, of course).

However, with the possibility for AI empowered quality translations to automatically make any (written) work accessible across any language, I wanted to at least raise some linguistic awareness. Ironically, I can see a day where these automatic translations could — depending on text and reader — be better adjusted for individual reading comprehension than the original. Well, let’s apply that thought-experiment here as well. AI prompt: “Rewrite this piece so that I can actually understand something.” A roundabout way for saying ELI5, so to speak.

The Translation Analysis

#

In this section, I want to move to the actual translations of this text, Crosslinguistic Musings. To that effect, I have created both a German and a French translation. Of course, in order to be able to make sense of each one of them you’d have to understand the languages yourself. If you do, feel free to check them out. And in case you don’t, just stick around for the analysis as I’m going to share a few impressions based on an example paragraph.

The Approach

#

In order to translate the original, I have used local AIs i.e. locally hosted large language models (LLMs). Here are my reasons:

• runs on my own computer vs. someone else’s (a.k.a the cloud)
• no internet connection necessary
• no subscription / APIs to deal with
• no usage limits
• private

Initially, I have tested two locally hosted LLMs, both released in early 2025. One is Gemma 27B / Q8 and the other one Mistral Small 24B / Q8. My hardware: 12GB VRAM / 64GB RAM. The final (and chosen) translations were then created with the Gemma model. Translation efforts were shared as follows:

• roughly 95% translated by local LLM
• remaining 5% adapted by myself

For this short analysis, I want to focus on the lack of perfection. Specifically, I’m going to single out a small paragraph from the original to give you a bit of an idea for the local LLMs’ challenges, how I would’ve approached the translation, and how I would’ve written the paragraph if it were the original instead (for the German version). And, finally, I’m going to compare the results with proprietary and/or cloud-hosted LLMs.

The focus is on the following translation sample:

I’ve been told by a few people I personally know that they’d love to read (some of) my work. Unfortunately, their English is almost non-existent. Well, they say unfortunately, but if you’re already familiar with any of my writings, I think we can both agree that the prefix un- can be dropped.

The Sample Analysis: German

#

In German, our translation sample poses one syntactic challenge and a wordplay that should easily be translatable. The prompt itself is kept simple (for the original in its entirety):

Please translate the following text to German...

And here is one of the local model’s ouputs:

Gemma 28B / Q8

Mir haben ein paar Leute, die ich persönlich kenne, gesagt, dass sie gerne (einen Teil) meiner Arbeiten lesen würden. Leider sind ihre Englischkenntnisse fast nicht vorhanden. Nun sagen sie “leider”, aber wenn du mit meinen Texten schon vertraut bist, denke ich, wir können uns beide darauf einigen, dass das Präfix un- weggelassen werden kann.

We’ll focus on the parts highlighted in bold:

1. syntax/grammar: […] that they’d love to read (some of) my work. → […] dass sie gerne (einen Teil) meiner Arbeit lesen würden.
2. wordplay/humor: Unfortunately…that the prefix un- can be dropped. → Leider…dass das Präfix „un-“ weggelassen werden kann.

Syntax / Grammar

Starting with syntax, in the English orginal the parenthesis (some of) can be removed without affecting the grammar:

[…] that they’d love to read (some of) my work.

The same isn’t true for the Gemma’s German translation

[…] dass sie gerne (einen Teil) meiner Arbeiten lesen würden.

Depending on whether we read the phrase with our without einen Teil, the grammatical case will change:

• dass sie gerne einen Teil meiner Arbeiten lesen würden → Genitiv
• dass sie gerne meine Arbeiten lesen würden → Akkusativ

That said, this entire construction wouldn’t really work for me. Sounds a bit off. Here’s how I would translate it myself:

Ein paar Bekannte haben mir gesagt, dass sie meine Arbeiten auch gerne lesen würden. Zumindest einige meiner Texte.

At this point, we’re somewhere between translation and adaptation (including changing the voice). As the author, however, I’d be cool with this solution — including favoring einige meiner Texte over einige davon in order to avoid a syntactic ambiguity where einige davon could refer both to Arbeiten or Bekannte. Yes, context should make it obvious that it could only be Arbeiten, especially because we already have ein paar (here a synonym for einige) in front of Bekannte. However, when you work in an excessive feedback culture for long enough…you get the idea.

Wordplay / Humor

When it comes to the morpheme-based wordplay un/fortunately, it should hopefully be self-evident that an LLM was behind the translation. After all, that’s the kind of mistake no human who is able to count to ten would ever make.

With that in mind, translating unfortunately with leider is absolutely correct. Sadly, to keep the humorous part dass das Präfix „un-“ weggelassen werden kann becomes somewhat nonsensical. Then again, maybe this is some form of AI avant-garde humor? Well, not on my watch, Gemma!

Now, the most obvious translation is a direct one. So here’s how I would translate it:

Unglücklicherweise sind ihre Englischkenntnisse kaum vorhanden. Gut, sie sagen “unglücklicherweise”, aber wer mit meinen Texten bereits vertraut ist, wird mir wohl zustimmen, dass die Vorsilbe “un-” hier getrost weggelassen werden kann.

The solution is to look for German synonyms for leider with the same morphology as the English unfortunately. And indeed, there is working equivalent: un/glücklicherweise. So we’re in luck.

Finally, would I be inclined to use the same pun in German if I were to write the text from scratch? The answer’s probably no. Unglücklicherweise isn’t an option that sounds appealing to me. Personal preference would gravitate towards leider as well (you’re forgiven, Gemma). Of course, that would make the morpheme-based wordplay a lot more difficult. Impossible even. However, there’s another synonym called bedauerlicherweise. I really like the sound of that one and its suffix -weise lends itself to a wordplay I’d enjoy in German:

Bedauerlicherweise sind ihre Englischkenntnisse kaum vorhanden. Wirklich “weise” wäre es allerdings, “bedauerlicher-” durch “glücklicher-” zu ersetzen.

Here, the pun is based on the suffix -weise (e.g. as in English likewise), which is then used as the adjective wise to state that it’d be wiser to swap bedauerlicher- with glücklicher-, retaining the core meaning behind the joke based on un/fortunately. The difference: to me, it’s slightly more sophisticated.

The Sample Analysis: French

#

In French, the sample appears to be better suited syntactically than the German version. At the same time, the difficulty regarding the wordplay remains, but it helps to clarify an ambiguity within the German translation. Let’s have a look:

Gemma 28B / Q8

Plusieurs personnes que je connais personnellement m’ont dit qu’elles aimeraient lire (une partie de) mes écrits. Malheureusement, leur anglais est presque inexistant. Enfin, ils disent “malheureusement”, mais si tu es déjà familier avec mes textes, je pense qu’on peut s’accorder sur le fait qu’on peut se passer du préfixe in- sans problème.

Syntax / Grammar

Just as in English and unlike in German, the parenthesis can easily be approached as optional:

[…] qu’elles aimeraient lire (une partie de) mes écrits.

Nothing to add here. It just works.

Wordplay / Humor

The pun is where it gets interesting. From a purely morphological perspective, malheureusement is excellent as it comes with the same structure as the English unfortunately: prefix + base + suffix. In other words, to recreate the wordplay from the original, all Gemma would have to do is to drop the prefix mal-: → heureusement, French for fortunately.

Instead, Gemma came up with this:

[…] qu’on peut se passer du préfixe in- sans problème.

Basically, the LLM failed to associate the mentioned préfixe to its referent malheureusement and, instead, translated un- (English) in isolation to in- (French).

This is particularly useful information when it comes to the German translation since it still contained the prefix un-, leading to the following possibilities:

• the prefix was not translated (unchanged from English)
• the prefix was translated (since the German equivalent is also “un-”)

Adding the French example into the mix, it becomes a lot more likely that the model simply translated the prefix in isolation. After all, a simple LLM cannot reason about language the way we humans do, right? So how is it supposed to “understand” that a linguistic component like un- in isolation actually refers back to another word like fortunately without being explicit in the text itself?

Here’s where it gets interesting. The Gemma 27B model tasked with the translation is not a reasoning model — leading us to the question: how well would a reasoning (or “thinking”) model do with this challenge?

The Reasoning Comparison

#

At the time of writing, it is May 2025. In other words, if you’re reading this sometime in the (not so distant) future, the following table might be more amusing than enlightening. Change is constant and rapid, after all. In the meantime, let’s have a look at a comparison between different reasoning models regarding the translation of the wordplay un/fortunately. As a reminder:

Unfortunately, their English is almost non-existent. Well, they say unfortunately, but if you’re already familiar with any of my writings, I think we can both agree that the prefix un- can be dropped.

Of note: I have used the free versions for all cloud-based LLMs. Prompt was kept the same:

Please translate the following text to German|Fench

Model Name	Local Possible	Locally Tested	Solution (German)	Solution (French)	Reasoning (Sample)
ChatGPT	No	–	Incorrect	Incorrect	Model thought output… (Comment gibru: ChatGPT’s thought process was hidden) German Translation: Naja, sie sagen ‚leider‘, aber wenn du schon mit irgendwelchen meiner Texte vertraut bist, können wir uns wohl darauf einigen, dass man das Präfix ‚un-‘ getrost weglassen kann.“ French Translation: Eh bien, ils disent “malheureusement”, mais si tu connais déjà certains de mes écrits, on peut tous deux convenir que le préfixe “un-” peut être supprimé.
DeepSeek (R1)	Yes	No	Incorrect	Correct	Model thought output… Reasoning for German: Here, “unfortunately” is “leider,” and the play on words is about removing the “un-” to make it “fortunately.” In German, “unfortunately” is “unglücklicherweise” or “leider,” but the prefix play would be on “un-” as in “un-” + “fortunately” (glücklicherweise). So the sentence would need to reflect that removing the “un-” from “unglücklicherweise” makes it “glücklicherweise.” German Translation: “Leider ist ihr Englisch so gut wie nicht vorhanden. Nun, sie sagen leider, aber wenn du bereits mit irgendeinem meiner Texte vertraut bist, denke ich, wir sind uns beide einig, dass die Vorsilbe »un-« getrost weggelassen werden kann.” (Comment gibru: Indeed, correct reasoning, wrong answer) Reasoning for French: The key challenge is the pun on “unfortunately” becoming “fortunately” by dropping the “un-”. In French, “unfortunately” is “malheureusement”, and “heureusement” is “fortunately”. So the prefix here would be “mal-”, so the joke would be removing “mal-” from “malheureusement” to get “heureusement”. French Translation: Malheureusement, leur anglais est quasi inexistant. Enfin, elles disent “malheureusement”, mais si tu connais déjà mes œuvres, je pense que nous sommes tous deux d’accord pour dire que le préfixe mal- peut être supprimé sans hésitation.
Gemini 2.5 Pro (exp)	No	–	Correct	Correct	Model thought output… (Comment: Was the only model to offer both the formal (Sie/vous) and informal (du/tu) translation without being prompted. However, it didn’t show a detailed enough thought process anymore. So I had to ask it to elaborate) Gemini’s answer: The core of the joke is the linguistic manipulation of “unfortunately” by suggesting the removal of “un-”. This kind of wordplay doesn’t always translate directly. I need to see if the target languages (German and French) have a similar prefix for their equivalent of “unfortunately” or if I need a different way to convey the same humorous idea. If the target language has a similar prefix that can be “removed” from its equivalent of “unfortunately” to create the same humorous effect, that’s ideal. If not, find another way to convey the self-deprecating idea that it’s “fortunate” they can’t read the work because it’s supposedly bad. This might involve a different turn of phrase that achieves the same sarcastic punchline. German Translation: Unglücklicherweise ist ihr Englisch so gut wie nicht vorhanden. Na ja, sie sagen unglücklicherweise, aber wenn Sie bereits mit meinen Schriften vertraut sind, denke ich, können wir uns beide darauf einigen, dass das Präfix “un-” weggelassen werden kann. French Translation: Malheureusement, leur anglais est quasiment inexistant. Enfin, ils disent malheureusement, mais si vous connaissez déjà mes écrits, je pense que nous pouvons tous les deux convenir que le préfixe “mal-” peut être supprimé.
Qwen2.5-Max	No	–	Correct	Correct	Model thought output… Reasoning for German: The original English relies on the word “unfortunately” containing the prefix “un-”, which when removed becomes “fortunately.” But in German, “leider” doesn’t have that prefix. So the joke is lost in translation. How can I handle this? Maybe the user is aware of this and still wants the translation, even if the pun doesn’t work. Alternatively, perhaps there’s a way to adapt the joke. Let me think. If the German sentence uses a word that does have the “un-” prefix, but I’m not sure. For example, “unglücklicherweise” means “unfortunately,” which does have the “un-” prefix. So if the original had used “unglücklicherweise” instead of “leider,” then dropping the “un-” would make “glücklicherweise” (fortunately). But the original uses “unfortunately,” which is more commonly translated as “leider” in German. However, to preserve the pun, maybe using “unglücklicherweise” would be better. Let me check that. So if the first sentence is “Unglücklicherweise ist ihr Englisch fast nicht vorhanden.” Then the second part: “Nun, sie sagen ‘unglücklicherweise’, aber wenn Sie meine Schriften kennen, können wir beide zustimmen, dass das Präfix ‘un-’ weggelassen werden kann.” Then removing “un-” from “unglücklicherweise” would result in “glücklicherweise,” which is “fortunately.” That way, the pun works. But the original uses “unfortunately” which is more naturally “leider” in German. However, “leider” doesn’t have the “un-” prefix, so the joke doesn’t translate. Therefore, to preserve the wordplay, it’s better to use “unglücklicherweise” even if it’s a bit more formal. German Translation: Unglücklicherweise sind ihre Englischkenntnisse fast nicht vorhanden. Nun, sie sagen ‚unglücklicherweise‘, aber wenn Sie bereits mit meinen Schriften vertraut sind, denke ich, können wir beide zustimmen, dass das Präfix ‚un-‘ weggelassen werden kann. Reasoning for French: In French, “Malheureusement, leur anglais est presque inexistant.” But the user’s original has a play on words with “unfortunately” and dropping the “un-” prefix. So in French, “malheureusement” can be split into “mal” and “heureusement”. The pun would be that “mal” can be dropped, leaving “heureusement”. French Translation: Malheureusement, leur anglais est presque inexistant. Eh bien, ils disent « malheureusement », mais si vous connaissez déjà certains de mes écrits, je pense que nous pouvons convenir tous les deux que le préfixe « mal- » peut être supprimé.
Qwen3:32b-q8.0	Yes	Yes	Incorrect	Incorrect	Model thought output… Reasoning for German: The first sentence is “Unfortunately, their English is almost non-existent.” The key words here are “unfortunately” and “almost non-existent.” In German, “unfortunately” is typically “Leider” or “Bedauerlicherweise.” “Almost non-existent” could be “fast nicht vorhanden” or “fast gar nicht vorhanden.” So the first part might translate to “Leider ist ihr Englisch fast nicht vorhanden.” Next part: “Well, they say ‘unfortunately’, but if you’re already familiar with any of my writings, I think we can both agree that the prefix ‘un-’ can be dropped.” The user is being a bit playful here, suggesting that the negative prefix isn’t needed. The word “un-” in German is often “un-” as well, but sometimes “un-” becomes “un-” or “un-” depending on the word. For example, “unbedeutend” (insignificant) has the “un-” prefix. So the phrase “the prefix ‘un-’ can be dropped” would be “die Vorsilbe ‘un-’ weglassen kann.” German Translation: Leider ist ihr Englisch fast nicht vorhanden. Nun, sie sagen zwar “leider”, aber wenn Sie bereits einige meiner Schriften kennen, können wir uns beide darauf einigen, dass die Vorsilbe “un-” weggelassen werden kann. Reasoning for French: First, translating “Unfortunately, their English is almost non-existent.” In French, “Unfortunately” is “Malheureusement”. “Their English” would be “leur anglais”. “Almost non-existent” – “presque inexistant”. So the first sentence becomes “Malheureusement, leur anglais est presque inexistant.” Next part: “Well, they say ‘unfortunately’, but if you’re already familiar with any of my writings, I think we can both agree that the prefix ‘un-’ can be dropped.” “they say ‘unfortunately’” – “ils disent ‘malheureusement’”. “But if you’re already familiar…” – “mais si vous êtes déjà familiers avec…” “any of my writings” – “l’un de mes écrits”. “I think we can both agree” – “je pense que nous pouvons tous les deux convenir”. “the prefix ‘un-’ can be dropped” – “que le préfixe ‘un-’ peut être supprimé.” French Translation: Malheureusement, leur anglais est presque inexistant. Ils disent « malheureusement », mais si vous êtes déjà familiers avec l’un de mes écrits, je pense que nous pouvons tous les deux convenir que le préfixe « un- » peut être supprimé.

Note for reader mode:
The "Reasoning Sample" column uses modal windows to display a larger amount of text.

The Discussion

#

While it is true that, in my testing, the free version of ChatGPT was not able to correctly translate the wordplay and that its reasoning was completely hidden, people interacting with a paid version of a more capable ChatGPT version might have a different experience than what is reflected in my test. In fact, a couple of days later and after experimenting some more with differently (and confusingly) named versions, ChatGPT got the French version right every once in a while. However, since all the results for the table were obtained on the same day, I won’t update them. Lesson: just make sure to always read the discussion for a complete picture 😉.

That being said, Gemini 2.5 Pro (exp) and Qwen2.5-Max could indicate that this translation challenge is already a solved problem when using the right model for the job. Specifically, compared to non-reasoning models such as Gemma 27B, they seem to handle context with much more subtlety, “understanding” that the prefix un- belongs to a previously mentioned word. Personally, I’m impressed. However, a more nuanced look at the results still paints a slightly different picture when it comes to trusting a fully automated translation.

In my tests, I got both Gemini 2.5 Pro (exp) and Qwen2.5-max to give me the wrong answer in German. Here’s an example for Qwen2.5-Max.

Qwen2.5-Max: Prompt without formatting

Please translate the following text to German: Unfortunately, their English is almost non-existent. Well, they say unfortunately, but if you’re already familiar with any of my writings, I think we can both agree that the prefix un- can be dropped.

Without using any kind of visual indicators such as apostrophes (e.g. “Unfortunately”, “un-”), the model pretty much consistently gave the wrong answer using Leider. At times, its thought process didn’t even show any considerations for translation alternatives such as unglücklicherweise.

Qwen2.5-Max: Prompt with added (subtle) formatting

Please translate the following text to German: Unfortunately, their English is almost non-existent. Well, they say "unfortunately", but if you’re already familiar with any of my writings, I think we can both agree that the prefix "un-" can be dropped.

Using quotation marks seems to help the model to better understand that there’s something requiring particular attention. So at first, it appears that the translation is correct. However, repeating the exercise enough times shows that it is still not fully reliable: occasionally, it doesn’t take unglücklicherweise into account, leading to wrong translations again using leider. Or, in one instance, it simply left unfortunately untranslated, reasoning that:

[…] if we keep the English word in quotes, then the reference to the prefix “un-” makes sense.

Seems like after repeating the exercise one too many times, our digital friend turns into some sort of smart-ass.

On top of that, Qwen2.5-Max’s reasoning shows something that could also be observed in DeepSeek R1’s thought process for the German version: the model can find the correct answer inside its thought process, but will still output the wrong final answer. Here’s an example (prompt including the quotation marks):

Qwen2.5-Max:

So the original translation used “leider,” which doesn’t have the “un-” prefix, so the joke is lost. Therefore, to preserve the wordplay, it’s better to use “unglücklicherweise” in the German translation.

So the corrected translation would be:

“Leider ist ihr Englisch fast nicht vorhanden. Nun ja, sie sagen ‘unglücklicherweise’, aber wenn du bereits mit meinen Schriften vertraut bist, können wir beide zustimmen, dass das Präfix ‘un-’ weggelassen werden kann.”

That way, the play on the prefix “un-” works because “unglücklicherweise” becomes “glücklicherweise” when the prefix is removed.

And the final answer:

Leider ist ihr Englisch fast nicht vorhanden. Nun ja, sie sagen „leider“, aber wenn du bereits mit meinen Schriften vertraut bist, können wir beide zustimmen, dass das Präfix „un-“ getrost weggelassen werden kann.

So what’s going on here? My guess: word frequency. Basically, the more “familiar” a model is with a given word, the more likely it will choose it over a less frequent one to express itself. This could lead to contradictions in terms of reasoning and conclusion as the solution for the wordplay in German is the less frequently used term — something that can actually be observed in Qwen2.5-Max’s reasoning sample from the table:

But the original uses “unfortunately,” which is more commonly translated as “leider” in German.

Using a corpus provided by the University of Leipzig, I have compared leider and unglücklicherweise to get a bit of a scientific feel instead of solely relying on intuition (and taking the LLM’s word for it). A quick look at the overall occurrences of each word in the entire corpus shows a massive difference: 58'821 (leider) vs. 257 (unglücklicherweise).

By contrast, malheureusement is the most frequent translation equivalent in French. This time simply relying on intuition, but fairly confident as no model ever opted for a possible synonym. And neither would I if I were to translate the text myself.

That said, I have repeated the translation five times with Gemini 2.5 Pro (exp), DeepSeek R1, and Qwen2.5-Max. Gemini got the French translation right 5/5. On the other hand, DeepSeek R1 and Qwen2.5-Max both 4/5, once leaving the English prefix un-…well, unchanged. In the case of DeepSeek R1, the reasoning was somewhat amusing as well:

DeepSeek R1:

I think this should be the accurate translation, preserving the wordplay as much as possible by referencing the English prefix “un-”, even though in French the prefix for “malheureusement” is different. The alternative would be to adapt the pun to French by changing the prefix to “mal-”, but that would alter the original reference. Since the user asked for a translation, not a localization, sticking to the original’s reference to the English prefix is better, even if it requires the reader to understand a bit of English.

Then, the model proceeds to use un- as the prefix instead of mal-.

All things considered, the models’ higher success rate with the French translation compared to the German one can very likely be explained by the fact that, unlike in German, the most frequent French translation equivalent for unfortunately also happens to be the one with the same morphology. At the same time, it is still impressive that it can actually reason when it comes to the connection between a prefix in isolation and the preceding word it belongs to — however (occasionally) flawed the final answer may be.

Lastly, my locally tested and far smaller reasoning models couldn’t correctly translate that wordplay, regardless of the formatting. Interestingly enough, however, just like its larger cousin Qwen2.5-Max, Qwen3:32b-q8.0 also explicitly refered to the frequency hypothesis in one of the tests:

Qwen3:32b-q8.0

Okay, let’s tackle this translation. The user wants the English text translated into German. First, I need to understand the original message thoroughly. The first sentence is straightforward: “Unfortunately, their English is almost non-existent.” The key here is to convey the regret or disappointment about their lack of English skills. In German, “Unfortunately” can be “Leider” or “Unglücklicherweise”. “Leider” is more common in everyday use, so I’ll go with that.

Moreover, it always left the prefix un- as is — including in French. In other words, the locally tested Qwen3:32b-q8.0 was simply not capable to apply any sort of reasoning to the prefix itself or to connect the dots between the prefix and the word it belongs to. This means that, even when being aware of the option unglücklicherweise as evidenced by the reasoning sample above, providing the correct answer, un+glücklicherweise, would simply be a stroke of luck rather than based on actual reasoning.

So one of the take-aways here is that, as of now, comparatively smaller models (that are easier to run locally) cannot compete with their much larger counterparts when it comes to solving this particular linguistic challenge.

As for the larger models, it might be interesting to look into the dynamics between reasoning capabilities and a preference for more frequent occurrences. More importantly, however, exposing a model’s reasoning or thought process seems incredibly valuable to get a better understanding for how to interact with it in a more meaningful way.

The Conclusion

#

Without proper context, it might seem that the translation analysis disproves my observation in the main part: how are AI empowered translations supposed to automatically make any (written) work accessible across any language if they cannot reliably translate something as simple as a basic pun?

At the time of writing, the answer is fairly straightforward: they are indeed not reliable enough — but they will be.

In order to add some context, a question might help: how have things been evolving over the past couple of years? Based on my own experiments, the answer is fast. Extremely fast, even. In late 2022, running an LLM locally was pretty much unthinkable from my perspective. I became aware of RWKV in early 2023 and with its 14B parameters it seemed like it would take years for me to be able to run an LLM of that size locally on my potato computer. Couple of months later I was already experimenting with larger models. Same potato.

Furthermore, quality improvements regarding the models’ output are no different. Since this piece is specifically about translation, let’s stay within that particular domain. One of the experiments I kept repeating was to give my local LLMs texts with challenging syntax. The challenging part was defined based on an observation that they simply couldn’t get the translation right. At the first. Then, newer and bigger models started to provide the correct translations and, soon after, even smaller ones were giving quality answers. Essentially, there’s no reason to think that something as trivial as a pun or any kind of syntax is going to pose a serious challenge for much longer. At the end of the day, however, time will tell, right?