ProZ.com Translation Forums

Survey: how MT & AI are impacting in the work of legal translators - your opinion counts

MARIA JOSE RUBIALES SILGUERO — Thu, 01 Oct 2026 15:22:00 +0000

Forum: Post-editing & Machine Translation
Topic: Survey: how MT & AI are impacting in the work of legal translators - your opinion counts
Poster: MARIA JOSE RUBIALES SILGUERO

Dear colleagues,

I am a doctoral researcher at the University of Geneva's Faculty of Translation and Interpreting (Transius Centre), researching how machine translation and AI are impacting legal translators in the private sector. The aim of the study is to better understand how these technologies are influencing professional practice and legal translators’ perceptions, with a view to producing evidence of the current challenges and added value of human expertise in the field.

If you work in legal translation (freelance or in-house) and have experience with MT/post-editing, I would love to hear your views! The survey is anonymous, takes about 15 minutes, and is available in English, French and Spanish.

Survey in English:
[url removed]

Survey in French:
[url removed]

Survey in Spanish:
[url removed]

If you have any questions about the study, feel free to reach out at [email removed] .

Thank you very much for your time, your input would really help.

Kind regards,
María José Rubiales

Appalling quality of "NMT" matches - have you experienced this? | Oops I confused slope and slop

Daryo — Thu, 01 Oct 2026 04:05:00 +0000

Forum: Post-editing & Machine Translation
Topic: Appalling quality of "NMT" matches - have you experienced this?
Poster: Daryo
Post title: Oops I confused slope and slop

I'm glad I have at least inadvertently provided some light entertainment.

Although as it’s often the case we laugh at something that should make us cry.

I dread to think about the number of unsuspecting readers who will quote the AI's confabulations (a.k.a. AI slop) as 'THE authoritative explanation'. The damage done by Google Translate pales into insignificance by comparison.

[quote]Joakim Braun wrote:

[quote]Daryo wrote:

(...) and then a long winded completely different explanation:

[url removed]

My guess is that the first two are correct and the third one is one very large helping of AI slope.

[/quote]

That's like a satire of AI - shiny, verbose, confident and wrong.
See [url removed] [/quote]

Thanks for the link, but my knowledge of German can be measured only in homeopatic doses.

Appalling quality of "NMT" matches - have you experienced this? | Satire

Joakim Braun — Wed, 30 Sep 2026 13:29:00 +0000

Forum: Post-editing & Machine Translation
Topic: Appalling quality of "NMT" matches - have you experienced this?
Poster: Joakim Braun
Post title: Satire

[quote]Daryo wrote:

(...) and then a long winded completely different explanation:

[url removed]

My guess is that the first two are correct and the third one is one very large helping of AI slope.

[/quote]

That's like a satire of AI - shiny, verbose, confident and wrong.
See [url removed]

Appalling quality of "NMT" matches - have you experienced this? | Sloppety slop

Anna Sarah Krämer — Wed, 30 Sep 2026 13:11:00 +0000

Forum: Post-editing & Machine Translation
Topic: Appalling quality of "NMT" matches - have you experienced this?
Poster: Anna Sarah Krämer
Post title: Sloppety slop

[quote]Daryo wrote:

A digression.

Or maybe not that much of a digression, as it might turn out to be a good illustration of the 'benefits' of shoddy use of AI.

...

The variation on 'go indulge in some sex and travel' is the right one - am I guessing right?
[/quote]

From the long winded explanation, which might not be correct, but is entertaining enough in its overexcited feverdream verbarrhea to have me snort out in disbelief:

[quote] ... Its literal translation may leave one perplexed; however, delving deeper into its figurative meaning unveils a fascinating insight into cultural nuances and linguistic richness.

Within this idiom lies an invitation to embark on an adventure beyond conventional boundaries, to explore uncharted territories where pepper plants thrive. By venturing where pepper grows abundantly, individuals are encouraged to step outside their comfort zones and embrace unfamiliar experiences. The essence of this idiom lies in embracing challenges head-on and seeking personal growth through traversing unexplored paths. ... [/quote]

Beautiful. It's a prime example of word slop. So many words! They sound good! They evoke emotions! Put together, they make up a text! It's good to fill your web page with lots of words!

Appalling quality of "NMT" matches - have you experienced this? | About peppers ...

Daryo — Wed, 30 Sep 2026 12:57:00 +0000

Forum: Post-editing & Machine Translation
Topic: Appalling quality of "NMT" matches - have you experienced this?
Poster: Daryo
Post title: About peppers ...

A digression.

Or maybe not that much of a digression, as it might turn out to be a good illustration of the 'benefits' of shoddy use of AI.

This recommendation to 'go where peppers grow' picked my curiosity.

Found three explanations:

hingehen, wo der Pfeffer wächst (idiomatic) to go jump in the lake
[url removed] ,_wo_der_Pfeffer_w%C3%A4chst

You usually say this to someone you don’t want to see or be near you, especially if you’re angry,
[url removed] #:~:text=You%20usually%20say%20this%20to%20someone%20you%20don%E2%80%99t%20want%20to%20see%20or%20be%20near%20you%2C%20especially%20if%20you%E2%80%99re%20angry%2C

and then a long winded completely different explanation:

[url removed]

My guess is that the first two are correct and the third one is one very large helping of AI slope.

The variation on 'go indulge in some sex and travel' is the right one - am I guessing right?

Appalling quality of "NMT" matches - have you experienced this? | Speak up

Anna Sarah Krämer — Wed, 30 Sep 2026 10:24:00 +0000

Forum: Post-editing & Machine Translation
Topic: Appalling quality of "NMT" matches - have you experienced this?
Poster: Anna Sarah Krämer
Post title: Speak up

[quote]Zea_Mays wrote:

You surely already did it, but it's paramount to inform clients about the bad quality in order to educate them about the lies tech bros are using to sell them their ai services. Not to mention the increasingly worrying data security aspect. [/quote]

I did that, and so far have gotten no answer. But I know if we keep insisting, more people will listen, so I will keep telling my clients. This boat is slowly turning around and businesses are starting to realize they've been scammed by the tech bros. It doesn't matter how 'future proof' you want your company to look like - some quality problems can't be hand-waved away.

One of the most ridiculous offers in the last days was a new AIPE scheme: get paid only for what you edit, get paid nearly nothing for what you just read. On top of the already laughably low MTPE rate. Since I see more work in my language pair appearing again I told them to 'go to where the pepper grows', as we say in German, and I'd encourage all colleagues who can afford it at the current moment to do the same.

Anybody else working for Propio

Erika Lundgren — Tue, 29 Sep 2026 18:53:00 +0000

Forum: Post-editing & Machine Translation
Topic: Anybody else working for Propio
Poster: Erika Lundgren

Hi! I have been working for ULG for years, and now since the acquisition for Propio, but the last month or so I have noticed a serious decline in jobs. Does anybody else have this experience? Wondering if they are struggling to keep clients.

Appalling quality of "NMT" matches - have you experienced this? | Tell the client

Zea_Mays — Tue, 29 Sep 2026 14:17:00 +0000

Forum: Post-editing & Machine Translation
Topic: Appalling quality of "NMT" matches - have you experienced this?
Poster: Zea_Mays
Post title: Tell the client

You surely already did it, but it's paramount to inform clients about the bad quality in order to educate them about the lies tech bros are using to sell them their ai services. Not to mention the increasingly worrying data security aspect.

Appalling quality of "NMT" matches - have you experienced this? | NMT is the classical MT we've been using for about 10 yrs

Li-Hsiang Hsu — Tue, 29 Sep 2026 14:00:00 +0000

Forum: Post-editing & Machine Translation
Topic: Appalling quality of "NMT" matches - have you experienced this?
Poster: Li-Hsiang Hsu
Post title: NMT is the classical MT we've been using for about 10 yrs

However, the situation has now evolved. Some providers can implement a routing system that allows switching between LLMs and NMT based on the complexity of a given task or the length of the context. They can also employ a hybrid workflow combining NMT and LLMs, depending on the architectural design. But they still call it NMT.

Appalling quality of "NMT" matches - have you experienced this? | No transparency about the products used

Anna Sarah Krämer — Tue, 29 Sep 2026 12:56:00 +0000

Forum: Post-editing & Machine Translation
Topic: Appalling quality of "NMT" matches - have you experienced this?
Poster: Anna Sarah Krämer
Post title: No transparency about the products used

[quote]Hideki Yoshida wrote:

I guess all MT systems today, including "LLM translation" or Transformer-based MT, use some kind of NMT (neural MT).
"Good old MT" reminds me of statistical MT, which has mostly become extinct. [/quote]

I believe one problem is that providers are not really transparent about the technology driving their products. AI is a marketing term, not a description. I also believe new startups without deep knowledge of the language industry and the intricacies of translation might be trying to stake their claim, in this case with ridiculous results.

My comparison has always been Deepl - they did have a short period of their product failing in the same way I see with these "NMT" segments, and I suspect they have backpedaled to whatever worked before their tentative introduction of "AI". Most of the time their EN-DE translations have been reliable enough to save time.

With agency clients trying to push prices further for AI vs MT I believe it's important to push back. In this case I have alerted the client to the bad quality, because it can't possibly be in anyone's interest to make translation more difficult while also making the output worse - and that's what is happening in this case.

[Edited at 2026-09-29 12:57 GMT]

Appalling quality of "NMT" matches - have you experienced this? | Definition of NMT

Hideki Yoshida — Tue, 29 Sep 2026 12:37:00 +0000

Forum: Post-editing & Machine Translation
Topic: Appalling quality of "NMT" matches - have you experienced this?
Poster: Hideki Yoshida
Post title: Definition of NMT

I guess all MT systems today, including "LLM translation" or Transformer-based MT, use some kind of NMT (neural MT).
"Good old MT" reminds me of statistical MT, which has mostly become extinct.

Appalling quality of "NMT" matches - have you experienced this?

Anna Sarah Krämer — Tue, 29 Sep 2026 11:31:00 +0000

Forum: Post-editing & Machine Translation
Topic: Appalling quality of "NMT" matches - have you experienced this?
Poster: Anna Sarah Krämer

I am currently working on a project for a leading tool manufacturer. Instead of the good old MT this project uses something called "NMT" - I suspect they ran the material through a chatbot with the prompt: Translate this material into German! Make no mistakes!

The result is laughable/scary (depending of your mood of the day). With MT in its current state I had predictable errors but most segments were usable, and a good MT paired with the right topic could get me to a good editing speed with reliable output.

Now there's a random, very confusing mix of complete bullshit, generously sprinkled all over the text: spelling errors, grammar blunders, literal translation, context errors, stuff you only imagined in feverish nightmares. The strangest thing is: the shitbot gets it right most of the time. Then, in some random segment, there's chaos. And of course, there's no consistency of formal/informal address, no consistency of terminology, no consistency between matching segments - I have to assume the shitbot hallucinates a new translation for every segment it finds.

Until now I was operating under the assumption that "AI" in translation was just the same old MT with a shiny new label, but it's worse than that, they are actually having chatbots handle translations.

Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations | Lara 3 - Launch presentation tomorrow July 30th

Tomás Cano Binder, BA, CT — Wed, 29 Jul 2026 15:03:00 +0000

Forum: Post-editing & Machine Translation
Topic: Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations
Poster: Tomás Cano Binder, BA, CT
Post title: Lara 3 - Launch presentation tomorrow July 30th

The people from Lara told me some time ago that there would be some new features soon, and as it happens, they are presenting them tomorrow July 30th.

I strongly recommend that you attend if you to MTPE in technical/scientific/specialised translation.

Thus far, Lara has made us far more efficient in our MTPE projects compared to Deepl Pro, and chances are that Lara could be a solution for you too.

Here is a link to info on the launch presentation:
[url removed]

Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations | Deepl on the way out...

Tomás Cano Binder, BA, CT — Fri, 19 Jun 2026 09:23:00 +0000

Forum: Post-editing & Machine Translation
Topic: Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations
Poster: Tomás Cano Binder, BA, CT
Post title: Deepl on the way out...

I thought I'd report my experiences with Lara as a substitute of Deepl Pro. With Deepl's Classic model disappearing no matter what, as confirmed (not with these words of course) by a Deepl representative when I emailed Deepl's support about the issues, finding a suitable substitute was a must for me.

Over a period of 20 days, I have used up my first 1.5 millon characters in Lara and have in fact stopped using Deepl completely. With extra characters charged at 20 cents per 100,000 characters, I am not too worried about consuming beyond my monthly supply.

My experiences with Lara:
- No omissions, not even in long sentences with long enumerations (Deepl frequently destroyed meaning)

- Respects most formatting (Bold is something that goes missing at times) and handles figures nicely. It even adds non-breaking spaces before units of measurement, something I have done manually for many years with Deepl. Such a luxury to have it done as it should be!

- Respects the glossary almost always (as long as the locale IDs you use in the CSV files you upload match the exact locales Lara identifies when you translate; if you use a different locale even within the same language, it does not find the terms)

- Is faster than Deepl: you get a 10,000-word file translated in about 5-6 seconds.

- Is far more reliable than Deepl: with Deepl, I had recurrent issues almost every week, with glossaries not being honoured, the server being down, the app doing strange things in my computer... I have installed Lara's webpage as an app in my Windows computer (using Brave), and it works very reliably all the time. No glitch in 20 days. Such a luxury here too!

- It translates very fast also interactively (in the web page, as opposed to translating a document).

- You can tweak the style (more faithful, more creative...) and can even tell it things to do in a prompt-like manner. With Deepl, there was one single style.

My only two issues with Lara:
- It is not necessarily consistent translating long documents: unless there is terminology involved, it can translate similar sentences with rather different approaches. This is normally undesired in my line of work.

- It does not have a keyboard shortcut so that I can use it even with translation tools that have no Lara Plug-in. I have to copy/paste text around, which is not that efficient.

Having a solid alternative such as Lara, chances are I will phase out Deepl completely in the next couple of weeks and cancel my subscription. It is sad because Deepl was such a swell concept, but one which has not shown improvement over the years... more on the contrary both in reliability of the Deepl Pro app and in translation quality.

I will be evaluating other tools over the next weeks. Share your experiences!

Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations | Thank you!

Tomás Cano Binder, BA, CT — Sun, 07 Jun 2026 11:13:00 +0000

Forum: Post-editing & Machine Translation
Topic: Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations
Poster: Tomás Cano Binder, BA, CT
Post title: Thank you!

[quote]CafeTran Trainer wrote:
I assume these non-breaking spaces are required in many languages.
You can handle them with a regular expression, either on a segment-by-segment basis when confirming each translation, or across the entire project before running QA.[/quote]
Yes, this is very much what we do daily in our projects, although once per each unit of measurement we find in the documents, as opposed to doing a global replacement in the way you suggest. While I see the power in the RegEx you posted, I think we prefer to do it on the spot to avoid adding non-breaking spaces where they should not be.

In any case, if the MT tool does not destroy the non-breaking spaces and even adds them in the correct places, it already saves very valuable time (Deepl does not seem to do either of them at the time of writing this post).

Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations | Non-breaking spaces

CafeTran Trainer — Sun, 07 Jun 2026 10:14:00 +0000

Forum: Post-editing & Machine Translation
Topic: Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations
Poster: CafeTran Trainer
Post title: Non-breaking spaces

[quote]Tomás Cano Binder, BA, CT wrote:

- Adds non-breaking spaces before units of measurement, as required in Spanish[/quote]

I assume these non-breaking spaces are required in many languages.

You can handle them with a regular expression, either on a segment-by-segment basis when confirming each translation, or across the entire project before running QA.

The replacement string would be something like $1·$2 (depending on the regular-expression syntax supported by your CAT tool).

I've inserted the regular expression as an image, but feel free to contact me if you'd like the text version.

Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations | Using glossaries?

Tomás Cano Binder, BA, CT — Sat, 06 Jun 2026 15:20:00 +0000

Forum: Post-editing & Machine Translation
Topic: Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations
Poster: Tomás Cano Binder, BA, CT
Post title: Using glossaries?

[quote]expressisverbis wrote:
From my experience working with PT-PT using DeepL, I increasingly notice PT-BR lexical choices appearing in the output for a couple of years, for example, “garçonete” (waitress) instead of the standard European Portuguese “empregada de mesa”.[/quote]
I think that, also in Deepl, you can benefit from adding these things to your own glossary in the tool. However, I am seeing that Lara takes my glossary into account in a much higher degree.

I will continue to work with Lara for some days more and, if it behaves well with my main text types and my glossaries... chances are that we will move on from Deepl completely since at least we will have one good alternative. We will however try other options we have shortlisted in the office, although I might need a couple of months to try them all. I will keep you guys posted if I found any other good option in my are of technical and scientific post-editing!

Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations | I haven't, but

expressisverbis — Sat, 06 Jun 2026 14:26:00 +0000

Forum: Post-editing & Machine Translation
Topic: Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations
Poster: expressisverbis
Post title: I haven't, but

[quote]Tomás Cano Binder, BA, CT wrote:

Have you tried Lara in your target language combinations? Thus far I have only tried my main combination (EN-ES) and will certainly try my second one (DE-ES) soon, and will gather more experiences in the weeks ahead with more document types.

Have you tried Lara? What was the result for you? [/quote]

I haven’t tried LARA yet, but I’ve read that it performs quite well for certain language pairs and domains.
From my experience working with PT-PT using DeepL, I increasingly notice PT-BR lexical choices appearing in the output for a couple of years, for example, “garçonete” (waitress) instead of the standard European Portuguese “empregada de mesa”.
This likely reflects the fact that many MT systems are still heavily influenced by PT-BR training data, which can affect regional register in PT-PT, especially in subtitling/translation contexts where consistency and space constraints are important.
That is why, as a PT-PT translator, I spend considerable time refining post-edited output.

Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations | Trying Lara...... very successfully!

Tomás Cano Binder, BA, CT — Sat, 06 Jun 2026 09:18:00 +0000

Forum: Post-editing & Machine Translation
Topic: Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations
Poster: Tomás Cano Binder, BA, CT
Post title: Trying Lara...... very successfully!

I think I will have to thank Deepl for removing the Classic model in the end...

What we technical/scientific translators needed from Deepl was not that difficult:
- Translate faithfully
- Honour our glossaries
- Do not destroy formatting/use correct target-language orthotypographic rules
- Do not destroy meaning

I reckon Deepl was not really created with these considerations that are key to us specialised translators. After using Deepl for a number of years, I thought we simply had to endure the extra work fixing many little things.

I have translated a 10K-word document with a paid subscription of Lara ( [url removed] , Team Edition) and the results are surprisingly good:

- It honours the glossaries almost always (as opposed to Deepl around 90% of the time)
- It preserves most formatting (bold/italic/underline) and most non-breaking spaces
- Adds non-breaking spaces before units of measurement, as required in Spanish
- It does not destroy or alter the meaning, and I have found no hallucination yet (when Deepl added/repeated/deleted pieces of long segments)

I am focusing more on the meaning and text quality and spend virtually no time fixing a high number of little issues as I have done with Deepl for so long.

Have you tried Lara in your target language combinations? Thus far I have only tried my main combination (EN-ES) and will certainly try my second one (DE-ES) soon, and will gather more experiences in the weeks ahead with more document types.

Have you tried Lara? What was the result for you?

Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations | Fluent errors vs. clunky errors: why the distinction matters in LQA

Rodolfo Abachi — Wed, 27 May 2026 15:29:00 +0000

Forum: Post-editing & Machine Translation
Topic: Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations
Poster: Rodolfo Abachi
Post title: Fluent errors vs. clunky errors: why the distinction matters in LQA

Lieven, I understand your point and to some degree agree that it is our job to catch errors. But I think the distinction is worth exploring a bit more.

When an NMT output is clunky or literal, the very form of the sentence signals that something is wrong. The reader of the draft instinctively slows down and double-checks the source. With a fluent-sounding LLM output, the sentence reads naturally, the terminology sounds plausible, and the cognitive load of the review shifts. You are no longer pattern-matching for obvious errors. You are evaluating meaning at a deeper level, segment by segment, which is significantly more demanding.

I am not arguing that it becomes impossible to detect errors. I am arguing that the cognitive cost of detection goes up, and that this cost is not being reflected in how MTPE rates are negotiated or how clients perceive the value of post-editing.

From a practical LQA standpoint, I have noticed that LLM-generated output increases the risk of false negatives in quality assessment: the reviewer approves a segment that sounds right but is subtly wrong in terminology or meaning. That is a real issue in high-stakes domains, and it is not a matter of carelessness.

I think both positions here are valid in different contexts. For general content or marketing copy, fluent output is mostly what you need and the stakes are lower. For technical, financial, or legal translation, the risk profile is quite different.

Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations | Lucky you!

Tomás Cano Binder, BA, CT — Thu, 21 May 2026 08:42:00 +0000

Forum: Post-editing & Machine Translation
Topic: Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations
Poster: Tomás Cano Binder, BA, CT
Post title: Lucky you!

[quote]Lieven Malaise wrote:
I'm working with the Next-Gen model since it has become available. For Dutch it's way better than the classic option. Combined with your own glossaries it provides the best of both worlds. [/quote]
Into Spanish, and in technical/scientific translation... it is a big step back unfortunately. Count your blessings!

Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations | No problem

Lieven Malaise — Thu, 21 May 2026 06:01:00 +0000

Forum: Post-editing & Machine Translation
Topic: Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations
Poster: Lieven Malaise
Post title: No problem

[quote]Li-Hsiang Hsu wrote:
LLMs don't respect glossaries;[/quote]

You can create your own glossaries in DeepL, so that problem is already solved. It respects them really well.

If I'm not mistaken the DeepL Next-Gen technology is LLM-based, but not like "unleashed" LLMs like the chatbots we know.

I'm working with the Next-Gen model since it has become available. For Dutch it's way better than the classic option. Combined with your own glossaries it provides the best of both worlds.

Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations | Lokalise: placeholder management

Li-Hsiang Hsu — Thu, 21 May 2026 05:06:00 +0000

Forum: Post-editing & Machine Translation
Topic: Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations
Poster: Li-Hsiang Hsu
Post title: Lokalise: placeholder management

Basically, what Lokalise does is all about "placeholder management", a kind of "soft constrain": detection, protection, and round-trip conversion of placeholders in translation files. Lokalise does not perform any grammatical infection, embedding, encoding or decoding tasks, which should be implanted on the model side with transformers - that is the real hard constrain.

[url removed]

Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations | Lokalise

Li-Hsiang Hsu — Thu, 21 May 2026 04:29:00 +0000

Forum: Post-editing & Machine Translation
Topic: Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations
Poster: Li-Hsiang Hsu
Post title: Lokalise

[quote]CafeTran Trainer wrote:

[quote]Li-Hsiang Hsu wrote:

Known technical solutions such as imposing deterministic "hard constraints" and "glossary/constraint recalls" work though not perfectly, and are complicated to implement. [/quote]

Could you please provide some references for this? [/quote]

You can consult the Lokalise's page. In practice, a deterministic layer is added to constrain the LLM output. This works more or less well depending on the models. Furthermore, this layer is more effective when implemented on the LLM provider side than on a simple CAT tool.

[url removed]

[Edited at 2026-05-21 04:31 GMT]

Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations | Would love to read more about this

CafeTran Trainer — Thu, 21 May 2026 02:49:00 +0000

Forum: Post-editing & Machine Translation
Topic: Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations
Poster: CafeTran Trainer
Post title: Would love to read more about this

[quote]Li-Hsiang Hsu wrote:

Known technical solutions such as imposing deterministic "hard constraints" and "glossary/constraint recalls" work though not perfectly, and are complicated to implement. [/quote]

Could you please provide some references for this?

Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations | We should let them know

Li-Hsiang Hsu — Tue, 19 May 2026 15:23:00 +0000

Forum: Post-editing & Machine Translation
Topic: Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations
Poster: Li-Hsiang Hsu
Post title: We should let them know

What a shame! NMT models are definitely better suited than LLMs for technical translations. Why change a tool that works? LLMs don't respect glossaries; this is a known difficulty. Known technical solutions such as imposing deterministic "hard constraints" and "glossary/constraint recalls" work though not perfectly, and are complicated to implement. So it's possible that the DeepL developers don't impose "hard constraints" on their LLM.

Aren't LLMs supposed to make our lives easier, reduce our work time, and increase our productivity? This whole "free translation" and "free glossary" story shows the opposite.

Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations | Can we please...

Tomás Cano Binder, BA, CT — Tue, 19 May 2026 14:31:00 +0000

Forum: Post-editing & Machine Translation
Topic: Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations
Poster: Tomás Cano Binder, BA, CT
Post title: Can we please...

[quote]Epameinondas Soufleros wrote:
It's not his argument; it's what a chatbot spat out. [/quote]
Please, guys... The matter is quite serious actually. Let's contribute with useful information and respectful opinions.

Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations | —

Epameinondas Soufleros — Tue, 19 May 2026 14:20:00 +0000

Forum: Post-editing & Machine Translation
Topic: Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations
Poster: Epameinondas Soufleros
Post title: —

—

[Edited at 2026-05-19 14:50 GMT]

Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations | The problem

Tomás Cano Binder, BA, CT — Tue, 19 May 2026 14:11:00 +0000

Forum: Post-editing & Machine Translation
Topic: Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations
Poster: Tomás Cano Binder, BA, CT
Post title: The problem

[quote]Lieven Malaise wrote:
[quote]Rodolfo Abachi wrote:
the Next-Gen model's biggest problem isn't obvious errors — it's confident, fluent-sounding mistranslations that slip past even careful reviewers.[/quote]
You're not the first one to use this argument, but I will never understand it. It's literally our job to detect errors. Only bad/careless reviewers could be fooled by fluent-sounding text. [/quote]
As a colleague described in this forum, the problem is that clients and MLV's believe that Deepl's Next-Gen and other LLM-based machine translations produce "better" translations, when in reality they produce "better sounding", but wrong translations. We need to educate our clients about the time losses this is causing and how to compensate them.

Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations | Disagree

Lieven Malaise — Tue, 19 May 2026 14:06:00 +0000

Forum: Post-editing & Machine Translation
Topic: Deepl Translator Next-Gen Model - Or how unfounded trust on an LLM destroys translations
Poster: Lieven Malaise
Post title: Disagree

[quote]Rodolfo Abachi wrote:
the Next-Gen model's biggest problem isn't obvious errors — it's confident, fluent-sounding mistranslations that slip past even careful reviewers.[/quote]

You're not the first one to use this argument, but I will never understand it. It's literally our job to detect errors. Only bad/careless reviewers could be fooled by fluent-sounding text.