Translate more than 10K words per day with AI

Some people claim they translate tens of thousands of words per day, thanks to the use of AI.

But how strong are those claims actually?

I don’t mean those people are lying. I mean “What do they actually call a translation?”

Recently I got a telling example of a job I did myself: 166 000 words in two days! I’m not kidding you. It means I did over 83 000 words PER DAY.

But here’s the thing: it were only 9 500 words to be checked, of course spread over several sentences (or translation units, as we call that in the translation business). In those source sentences words originally written in caps were replaced by tags. The only thing I had to do, was replacing the words in the translation by those tags.

That means the 166 000 words had already been translated. Only some of them had to be replaced.

I needed 2 days for that job, which means I did almost 5000 words per day. The tricky part was the tags had to end up in the right position and they had to fit in the grammatical structure. Because of that, the translation sometimes had to be rephrased.

It shows that claims of tens of thousands of words translated per day don’t tell us anything if we don’t get a detailed analyses of the work. How much if the original words have to be translated? How much can be skipped entirely? How many translation units only demand a partial translation or retranslation?

We have to be careful to take bold claims about huge translation outputs at face value, because anyhow, a human translator is not able to do much more than 2500 or 3000 words per day from scratch, with or without AI, or MT as we call it in the translation business.

AI zonder review

Het Tudor Pro Cycling Team (PRT) lijkt volgens een website enkele nieuwe leden te tellen, het gevolg van een AI-vertaling zonder controle. Om dat soort vergissingen te vermijden is er AI review nodig: iemand moet de vertaling nakijken.

Dat is op zich niets nieuws. Tot nu toe noemden we AI review gewoon MTPE: machine translation, post-editing. Machine translation of machinevertaling is de term die in de vertaalsector wordt gebruikt voor automatisch vertalen, een techniek die zich al meer dan tien jaar geleidelijk ontwikkelt, en net als AI steunt op enorme voorraden tekst.

In het geval van MT worden echter geen willekeurige bronnen gebruikt, maar vertalingen die al werden gemaakt voor een bepaalde klant. De nauwkeurigheid is dan ook groter dan als met gewone AI, zoals ChatGPT of Claude, wordt gewerkt, omdat alles op de eindklant is gericht. Die MT wordt geleidelijk ook bijgewerkt door intelligentere softwareroutines in te voeren. MT is dus niet vertalen met AI avant la lettre, maar de echte vorm van vertalen met AI.

Die jarenlange ervaring met MT heeft snel doen inzien dat nakijken door een menselijke vertaler, post-editing, onontbeerlijk is. En dat is precies de fout die velen maken als het om AI zelf gaat: ze gooien een tekst in een gratis tool die ze op internet hebben gevonden, of ze nemen de goedkoopste versie van een of andere aanbieder die beweert alles met AI te kunnen klaarspelen, en het nakijken laten ze achterwege. Het wordt AI zonder review.

Het resultaat wordt dan iets zoals op de foto: Michael Storer wordt Michael Magazijnmedewerker, Will Barta wordt Zal Barta, Florian Stork wordt Florian Ooievaar (note bene een Duitser) en Larry Warbasse wordt zelfs Larry Oorlogslager.

Wie ons niet gelooft, kan hier een kijkje nemen op de site van het team.

AI Expo 2026

It will take a year or two before we finally grasp what AI really is capable of, and what not.

Nobody has to doubt that the current claims are exaggerated, but in the meantime we try to be as good informed as possible.

As I have written earlier, here and elsewhere, as a translator we all use AI in some form since some years. I don’t know when it exactly started. It’s difficult to pinpoint the moment, because it was a gradual evolution in the way our software became smarter in the way it combined terminology databases and previous translations.

The word ‘intelligent’ is not, however, the way I would describe it. The software, also the so-called Artificial Intelligence, gives sometimes unexpected results, but always shows a lack of intelligence, understanding and logic. It only does seem so superfluously.

During the AI Expo 2026, organized by Proz.com, some speakers gave some insightful talks showing how AI lacks in-depth knowledge and mastering of language and translation.

Some translation managers don’t have a clue

Lately an agency asked me to check translations, and they wanted to pay me per “hits”.

A “hit” is a text segment that contains a number of words.

However, they can’t tell me how many words there will be per hit, nor can they tell me how many words there will be in total and how much the payment in total would be.

Not only is the rate per hit what you normally get per word, but you don’t know whether it’s for 1 word, 2, or maybe even 7 or 10.

The translator will have to judge the accuracy of some translated sentences and grade them based on how much the meaning of the source is transferred to the translation. It was for around 5000 hits, which could be anything from 5.000 to 50.000 words.

The contact person wrote me a hit “usually consist of short sentences”. So, the final result will be closer to 50.000 words than 5.000 words.

I have, of course, declined the offer, on the grounds that a) the price was too low, and b) it’s impossible to know the workload.

a) Too low, because this kind of work is usually underestimated by the managers. It takes much more time than expected once mistakes and changes crop up.

b) It’s impossible to know how much hours I would spend on it. The minimum of 5.000 words means at least 10.000 words have to be read, which is approximately 40 fully printed pages. If there is nothing to be changed, even checking whether it’s correct is going to take at least two hours. But if it turns out to be 50.000 words, and things have to be changed, it could mean at least 2 days of work. That means it’s impossible to plan such work.

If they don’t come up with a serious price per word instead of per hit, and if they don’t tell me the amount of words, this is impossible to plan. Whether you are allowed to work at your own pace or not, doesn’t matter. The pay is far too low anyway. Moreover, accepting such offers leads to unfair competition with better paying clients.

Still going strong for long-standing client

It is so long ago, that I barely remember when I started translating for them. ‘Them’ is a huge company manufacturing heavy building machinery, like excavators and tools such as tilt-rotators.
According to my archives it all began in 2009. In those days I translated off-line, although I had already been working on-line on tractor manuals.
The translation memories became more important, and the system used them more intelligently. Someday off-line changed to on-line for that company too. I helped them making the program a bit more clever, and nowadays we call it MPTE: machine translation – proof editing, the terminology we use in the translation industry for working with artificial intelligence.
The documents are safety instructions, user manuals, maintenance instructions, montage and demontage instructions, and so on.
Apart from the growing involvement of MPTE, there was also the increase of software used in the machines. Luckily I once was an apprentice Cobol programmer/analyst, and I had some interest in computers and programming. As a matter of fact, I build my own website around 2000, when blogs didn’t exist yet. Unfortunately that website is defunct nowadays. But my ICT background still helps a lot, especially because car technology without software has become unthinkable.
More than 15 years is indeed a longstanding client. That record won’t be broken easily.

building #machinery #excavator #tiltrotator #tractor #safety #instructions #manuals #maintenance #montage #demontage

Onverwacht vermoeiende MPTE

Maandag nam ik met plezier een MTPE-project aan voor een merk van luxe-auto’s.


Nu, dat is geen nieuws, ik neem opdrachten altijd met plezier aan.


Maar het bleek twee dagen behoorlijk doorwerken te zijn. Ik was een beetje verrast dat ik me na de eerste dag al tamelijk leeg voelde.


Zoals gewoonlijk met dit soort MTPE-opdrachten vraag je je af waarom het zo zwaar voelt, want er was niet veel te vertalen en de meeste woordenschatproblemen waren al aangepakt in de TB.


Ten eerste was het echter een enorme hoeveelheid woorden: meer dan 166 000! Gelukkig moest ik er zelf maar ongeveer 10 000 controleren.


Maar, hola: de AI had geen rekening gehouden met de TB! Dus zadelde de kwaliteitscontrole achteraf me nog met een hoop te controleren woordenschat op.


En een andere programma – misschien ook de AI – had alle woorden geschreven in hoofdletters vervangen door labels. Dat was interessant voor combinaties zoals “the STOP button”, maar het verving bijvoorbeeld ook “OFF” en “ON” in werkwoordgroepen (phrasal verbs) door labels.


Dat betekende dat “Tun ON” en “Turn OFF” er precies hetzelfde uitzagen. Het werd allemaal “Turn Label1” omdat de labels in elke afzonderlijke zin vanaf 1 werden genummerd.


Maar de vertaling “Zet Label1” zou dan “Zet OFF” of “Zet ON” worden, wat je niet echt Nederlands kunt noemen.


Elke vindplaats van een phrasal verb moest worden gecontroleerd in het document met de brontekst om zeker te zijn dat de vertaling in orde was.


Het zijn zulke zaken waardoor MTPE tijdroverender is dan verwacht, en de deadline halen je nogal wat kopzorgen kan bezorgen.


Maar, niet getreurd, het is gelukt. Het maakt echter wel duidelijk waarop MTPE soms onverwacht vermoeiend is.

Unexpectedly tiring MTPE

Monday I gladly excepted an MTPE project for a luxury car brand.


Well, I always gladly except a job, there’s nothing new to that.


But it turned out to be two days of hard work. I was a bit surprised that I felt a bit empty after the first day.


As usual with this kind of MTPE jobs you wonder why it does feel so hard, because there wasn’t a lot to translate and the vocabulary problems had mostly been taken care off.


However, first of all it was an anormous amount of words: more than 166 000! Luckily, I only had to take a look at approximately 10 000.


But, oh boy: the AI didn’t take the TB into account! All of that had to be checked afterwards.


And one or other program – maybe the AI too – replaced all words written in capitals by tags. That was very interesting for combinations like “the STOP button”, but it replaced “OFF” and “ON” from phrasal verbs by tags too.


That meant “Tun ON” and “Turn OFF” looked exactly the same. It all became “Turn TAG1”.


But the translation “Zet TAG1” would have ended up as “Zet OFF” or “Zet ON”, which can hardly be called Dutch.
Every occurrence of phrasal verbs had to be checked to be make it was translated OK. Things like that make MTPE more time-consuming than expected, and meeting the deadline can be a headache.


I pulled it off, but it became clear why MTPE can feel so unexpectedly tiring.

AI-modus van Google (4)

Ook deze vond ik eigenlijk wel grappig:

Peter Motte
(Geraardsbergen, 31 maart 1966) is een veelzijdige Belgische vertaler, auteur en publicist die vooral actief is in de werelden van sciencefiction, fantasy en strips.
Professionele Activiteiten

Vertaler & Ondernemer: Sinds 1997 runt hij Vertaalbureau Motte in Geraardsbergen, waarbij hij technisch vertaalwerk (automobiel, ICT) combineert met literair werk en vertalingen van bekende manga’s (o.a. Death Note, Bleach) voor uitgeverij Kana.
Literair Werk: Naast 13 jaar redactie van het tijdschrift De Tijdlijn, schreef hij poëziebundels en stelde hij de verhalenbundel Atlas (2013) samen.

Online Aanwezigheid
Peter Motte blogt over taal, literatuur en AI, en vermijdt sociale media.

“vermijdt sociale media”… Nou ja, eigenlijk is dat wel waar, maar het staat er wel nogal apodictisch.