May 20, 2023

Taking the trash out of trashlation in translation

Let’s be real, people. We need to start using the word trashlation to describe what too often passes for professional activity in the language services sector these days.

I’m looking at all of you DeepL fans among others. 🧐  And especially the LSPs who try to sell that garbage or other post-edited spew as translation. Such nonsense is seldom fit for purpose, and when I am not bent over the Porcelain God in “prayer”, I am always reminded of Mark Twain’s comment that the difference between something “close” to the right word and the right word itself is akin to the difference between a lightning bug and lightning. I would suggest that in some cases the consequential difference could be compared to being strapped to the electric chair for execution as opposed to watching the Natural magnificence of a lightning storm.
Recently, a German publisher sent me a very interesting file containing a table with the original German text in the first column, a post-edited DeepL result by a German who has lived abroad for some decades in the second column, in a third the post-edited result from an American living in Germany, and in the fourth column the results of some beta program called DeepL Write, yet another iteration of all of this AI trash the fanboys and -girls all declare will replace human writing efforts. Although I prefer to be left out of any discussions whatsoever involving the linguistic quality of MT output, the submitter seemed pleased by what he saw and insisted on having my opinion. So I gave it to him. Unfiltered. 😃  I suggested leaving me alone with that nonsense, and said I considered all of it to be hopeless trash. Actually, I might have been harsher than that, as I don't consider that sort of “writing“ acceptable or in any way fit for purpose, any purpose with which I'll be associated. It is physically painful for me to look at verbal garbage like that. I would rather stick my nose in a garbage can full of rotting meat and inhale deeply. The original German text was not bad, and it deserved real translation like someone I recommended could provide for that subject matter.

In a long consulting gig in the second half of 2022, which ended only in April of this year, I had a very close look at how badly the commercial models of translation service offered by most agencies are broken, badly, badly broken. Years of social engineering propaganda by unscrupulous promotors of machine translation and artificial "intelligence" have skewed expectations badly so that if the buyer is lucky the "best" service might be "good enough", though barely.

Project managers at agencies tell me of their soul-crushing duty to force ever lower rates on those external providers of translation services who typically do the real work sold as the agency's deceptive product. All the while, charts and graphs and other "quality metrics" tell the fairy story of superior delivery.

It's time for individuals providing independent services to take a different approach. Let the linguistic sausage providers (aka LSPs) eat their own product. Take the trash out of trashlation.

Of course there has been a lot of talk for years about the need to find more direct clients. Most conferences for translators have at least one presentation on this topic. But many of the recommended practices are dated or have become less effective as LSPs providing trashlation have increasingly gamed the search algorithms to make their pages appear to be those of independent individual providers, and then once there the buyer is treated to lies and distortions suggesting that they may be better off with the superior "full service" of the agency. Very few of the claims on such pages reflect actual practice, something I know very well as an insider providing technical assistance to companies for a very long time.

I have also read suggestions recently that search engine optimization (SEO) strategies may soon be a dead letter. Why? Companies developing search engines are rushing to implement large language model (LLM) functionality, such as that found in ChatGPT, any many expect that to blow the algorithm gaming strategy to Hell, negating much of the investments companies and individuals have made to increase the visibility of their web sites and the services offered there.

I don't know, really, what disintermediation strategies might be most effective these days, but at the very least individual traders should examine alternative representation strategies. I saw an interesting one recently on Fiverr, a platform I remembered only for the idiotic idea that every service should cost $5. Well, that has apparently changed.

Although their presentations were often far from a perfect, translators offering services on those platforms are able to structure the "gigs" in such a way that buyers can easily specify relevant conditions, such as project scope, urgency, etc. And service providers can avoid overbooking by applying various kinds of throttles based on order volume. Extra services, such as multiple revisions, project glossaries and many other extras I have marketing over the years are a snap to set up.

In fact, I was so impressed by what I saw that I am considering to create special services for some of the routine technical services I provide for translation workflow training videos, custom import filters, regex tools to translation and text QA, etc. Frequently, more of my time is spent gathering information on a client's requirements than I actually spend providing the implementable result. The flexible FAQ functions, intake questionnaire, portfolio and communication tools look like they can be a huge time-saver.

Compared to what I saw on Fiverr and a few similar, huge volume platforms, the structure of platforms like ProZ-dot-com and Translator's Café definitely look "last century". The times they are a-changin' and maybe we should be too in the platforms we use to promote services that our potential clients need.


  1. Recent metrics on a 2,000 word text I translated show that DeepL text (if you count an insertion and a deletion as one change) requires one change every 6.9 words. In this particular case, DeepL was right about 14.49% of the time. Not very impressive. What DeepL got wrong, it got hopelessly wrong. This result is in line with many similar experiments I have done.

    1. Allison, aside from the trashlation quality of DeepL results, it makes an absolute dog's breakfast of information for dates, time and currency, producing more formats for each than I would ever have imagined and requiring quite a few regex-based QA tools to be used to clean up the mess.

  2. Dear Kevin,

    I recently discovered Translation Tribulations and am delighted to see there are other translators who share the frustrations I have been experiencing daily throughout my career. Despite formal education as a translator, I did not realize what makes a good translation until I started working for my current employer. Unfortunately, I had to achieve this epiphany on my own. Sadly, it is one that most of my coworkers have not been blessed with. Simple techniques like turning German passive voice into English active voice elude them like a jackalope. Nevertheless, most of our clients have for years been satisfied with translations that could more accurately be called “oversettings”, probably because they don’t know any better.

    At the start of this year, the founder of the company I work for—obviously fed up with the clients constantly pushing our prices down—decided to sell his 30-year-old business to a company that is much younger but making waves in the industry with its novel AI-based approach. This is, of course, going to make translation faster and more cost-effective (read “cheaper”) for customers.

    Our new parent company gave us a chance to test their system. Without human intervention in the process, the translation I got back would not have been usable. However, their full service with preprocessing, quality assessment, and post-editing was much more up to snuff. Nevertheless, that still makes the human element potentially the weakest link in the chain. Ultimately, the quality of a translation is closely tied to the quality of the translator training. Many translation techniques and translations themselves—plenty of them highly debatable—at our company come straight from the former big cheese himself and have been passed down from one translation memory to the next.

    In all honesty, translation quality comes down to the quality of the source text and the deadline it is translated under. When a client sends a 30,000-word document they spent months creating and expects the translation back within a week, we should not be surprised to see a target text resembling a product of Google Translate ca. 2007. That is especially true when the source text is so incoherent and full of redundancies that it is itself akin to rotting meat.

    Two of our long-term clients recently resorted to using MT and AI systems of their own. The first, a major Swiss bank, now pretranslates its texts for us with an MT algorithm and then complains when the quality of the final text they get back does not meet their standards. The other client, a German state-level bank, has withdrawn most of its weekly investing newsletter, reportedly to be translated using an AI tool. Oddly, enough it still sends us the one section of that product that an AI system ought to be able to handle because of the repetitive structure of the content. The parts that always vary from week to week and frequently represent the author’s attempt at creative writing with newly invented metaphors and sentences of 30 words or more are what the client expects AI to handle. How do they come up with those decisions?

    Eventually, what the translation industry is going to wind up with is source texts written with the aid of an AI tool because the corporate clients most translation agencies serve will want those materials produced faster. Then the translators, especially ones at big agencies, are going to be translating with the aid of an AI tool because the deadlines and prices force them to resort to more advanced CAT tools. What is that old acronym in computerese: GIGO = garbage in, garbage out?

    1. Hi Kevin, sorry for the doublet (the first one was anonymous)
      I love your term "trashlation"

      The problem is that technical writing (and medical writing that impacts on my job) are also both made by poor UK/US writers (likely Chinese etc), and I suspect by chatgpt and other similar trash-tools, so the picture is even worse in my opinion YUK

      And I noted another issue: (too) many "peers" use since years Internet over dictionaries and other good/official references because they havethe results faster, and they usually consider "more reliable" the more frequent result

      Well, I am noting that the websites best indexed by Google are rubbish sites, with improbable names like, that are likely made by robots looking at the copy-pasted rubbish inside, so just picture what can jump out from those "pro-level searches"

    2. Garbage, garbage, every where,
      And all the text did stink;
      Garbage, garbage, every where,
      Nor anyone would think.

  3. Hi Kevin, I love this term "trashlation"
    The problem is that technical writing (and medical writing that impacts on my job) are also both made by poor UK/US writers (likely Chinese etc), and by chatgpt and other similar trash-tools, so the picture is even worse in my opinion YUK

    BTW, I noted another issue: (too) many "peers" use Internet over dictionaries and other good references because they have results faster, and they usually consider "more reliable" the more frequent result
    Well, I am noting that the 1st and more common website indexed by Google are rubbish sites, with improbable names like, that are likely made with a robot looking at the copy-pasted rubbish inside, so just picture what can jump out from those "professional searches"


Notice to spammers: your locations are being traced and fed to the recreational target list for my new line of chemical weapon drones :-)