2. Mai 2019

YouTube systematically manipulates subtitle translations into German, turns right-wing into right-wing extreme


(source)

A few days ago, I had an English language article translated into German to use it for my blog. When I was going trough the text checking for errors, I wondered why the German term for "right-wing extreme" was mentioned so often. As it turned out, the translation tool changed most attribution terms affiliated with being politically right-wing into right-wing extreme. Realizing that I have stumbled over an anomaly that should not be there, I decided to test other texts and different translation tools – most with the same result. Finally, I also took a look YouTubes automatic translation of subtitles and the results are so damning that it may break the entire company.



Counting keywords and looking for a possible translation bias between left and right



My goal was to find out two things. The first being whether political attributes like "far-left", "left-wing" or their counterparts of "far-right" and "right-wing" are being translated appropriately into German. The second and more important goal was to find out whether the political attributs assoicated with left or right are translated synonymously.

The basic premise for this approach is that "left-wing" and "right-wing" basically mean the same, but each for the other side of the political spectrum, where they represent areas in the political spectrum that are equally far away from the center. If that was not the case, it would indicate deliberate manipulations from the side of YouTube/Google to create a specific framing for their German speaking audience who rely on the translation function for subtitles.

As base for this analysis, I used Tim Pools commentary channel, where I went through 50 videos (each 10-15 minutes long) he created in the past two weeks and looked up how YouTube translated certain keywords. My main focus was on the terms "far-left", "left-wing","left-wing" and "right-wing" but I also counted a couple of other terms with political left/right attributes.

In total, I found 295 keywords of which 169 refer to the political left and 126 pointing to the political right. The four keywords mentioned above make up around 90% of the total lot, which is enough to make the results statistically stable. This of course provided I did a good job skimming through Tims rants.

Besides the human error dimension, it is also important to mention that the translation tool did not properly translate all keywords. Some were put into a different non-political context. Another problem is that the translations are not fixed but are translated live while the video plays. This means that if you want to check my list of keywords it may be that you get different results, since these automatic translations work with probabilities by which different synonyms are used. Therefore, you should check at least 20 different instances and count the results. All that matters in this regard is the statistical distribution of the different words used in the translation.

You can find the list with all the keywords I found here as an Excel/ODS file.

1) Keywords containing leaning, radical, center und liberal


The following table contains the set of political keywords which were used less often and which were not my main focus. The fact that they are turning up less often in Tims videos than the ones with "far-" and "wing-" does not mean they are less often used in general or less manipulated with, but only that Tim is using these terms less often in his political commentaries.


Most of the terms on both sides of the spectrum are being translated accordingly to their meaning, while especially the ones for the political right are accurate. On the left side though there are a few hints which may be interpreted as manipulation.

One example is that "far-left leaning" has been translated into the German far more often as "left-leaning" than into its literal equivalent, which effectively moves the attribute more to the center in the German version. The second indicator ist that almost half of all "left-leanings" - which is a centrist term - are turned into "left" which in its meaning has a clear political affiliation with the left side of the political spectrum and not the center as it should be.

What you can derive from this is that YouTube translations tend to move attributes from the far end of the politically left spectrum further to the center, while attributes that are associated more with the center are being moved to the left. Although the numbers are too low for definite statements, both observed biases may be interpreted as intention to create a certain political framing benefitting the left side of the political spectrum.

On a practical level, it means that when Tim Pool calls himself "politically left-leaning", almost half of his German audience is told that he is "politically left". That is quite a noticable difference.

2) Comparing the German translations for "far-left" and "far-right"


In the German language, the two terms "far-left" and "far-right" do not exist as political terms in their literal translation. Nonetheless, when you translate them into German, you can put them into a political context and assign them the same meaning as they have in English. Since this is very simple gramatically, no translation tool should have any problems with creating the correct context.


While the statistical analysis from further above only resulted in weak indications, this one is statistically by far more stable and it is very telling given the distribution of the different terms used in the German translation.

The bias though only occurs with the attributes on the right side of the political spectrum. In fact, most attributes on the left are quite accurate, of which the ones with "weit", "ganz" and "äußerst" - which make up 80% in total - hit the definition behind the political "far-left" pretty much on the spot. To my surprise, even the German version of "left-(wing) extremist" made it on the list, although I do not think the term is fits into the definition space for the "far-" attributes.

On the other side, unfortunately things look very different. "Right-(wing) extremist" is dominating the table with almost half of all counted attributes and turns up more than 10 times as often as its counterpart "left-extremist".

Equally disturbing is the direction on the left-right-spectrum for the respective translations. "Far-left" is translated five times more often in a way that it seems more centrist than extremist (15:3), whereas "far-right" is translated four times more often in a more extremist way than in a centrist (23:6). The result of that bias into different directions is a maximization of the gap of meaning between the two sides of the political spectrum. The bottom-line ration is around 4:1 in favor of politically left attributes.

On average, German translations of the term "far-left" lead to a slightly more centrist presentation of politically left contexts. Translations of the term "far-right" into German though move politically right contexts significantly further to the right - and with that not only further away from the center where they are located politically in the English version, but they are also framed to a massive degree as right-wing-extremist.

The conclusion is that YouTube/Google wants its Germans viewers to perceive "far-left" content as less extreme than it is, and it puts almost have of all "far-right" content right next to Hitler.


3) Comparing the German translations for "left-wing" and "right-wing"



In difference to the "far-" terms, "left-/right-wing" are equally established and defined in the German language as they are in English.



The statistical distribution shown in the table speaks for itself. At least I do not know how to explain this in any other way than deliberate manipulation to move content associated with the political right-wing to the extreme end of the spectrum in order to make it look bad. These 16.4% where "right-wing" has been turned into "rechtsextrem" are equal to almost one in six! Practically, this means that with its translation tool, YouTube makes German viewers believe that about every sixth aspect associated with right-wing is something extremist.

At the same moment, none of the translations for "left-wing" are moved to the extreme end of possible translations. I am not opposed to that - in the fewest cases something poltically right-wing or left-wing has anthing to do with extremism - but this bias between left and right is staggering given the apparent brazenness on the side of YouTube/Google to fool its customers in such a way.

You always have to consider the practical implications of this bias. In Tims videos there are several very telling examples that I went through. Take this video for instance at 3:35 where Tim says (or cites): "Far-right Tucker Carlson is angry." YouTube turns the sentence for its German viewers into this: "Der rechtsextreme Tucker Carlson ist wütend."


Screenshot from the linked video

Now, the name Tucker Carlson may not be very well known in Germany and I do not know how many Germans watch Tim Pool videos besides me. But for someone who stumbles over the video by chance - take a pupil for instance who is researching for a school presentation - and needs the translated subtitles because his English is not the best, that person will possibly think of Tucker Carlson as an extremist.

And Carlson is not the only one. It is the same problem with all sorts of political viewpoints which may be right-wing but is not extremist. Something like low taxes or gun ownership for instance. Thanks to YouTube, Germans have a good chance to be made believe that this is considered extremist in the English language world, even though it is clearly not.



Implications and political and legal consequences



The implications from this are huge given the fact that with the Internet, English as the global lingua franca is more important than ever, and many Germans despite their deficiancies when it comes to speaking English try to inform themselves about other countries using that language. Just alone in former Communist Eastern-Germany, where Russian was taught as first language instead of English, there are at least 10 million people aged 40+ who have to fully rely on the integrity of automatic translation services like the one provided by YouTube. As it appears they are being fooled by the company in the most fundamental way possible - not much differently to the ways in which they were fooled by their former Communist leaders.

With this kind of political bias YouTube/Google has basically institutionalized "fake news". Because even if the news is accurate in English, the news will be distorted and "faked" as soon as it goes through the translation process while containing the wrong keywords.

Still, do not expect too many or any consequences at all from the German government. They directly profit from this kind of faking and the legal system has a different modus when it comes to consumer fraud. Huge fines like in the United States do not happen in Germany. Slander of public figures is a punishable crime, but the fines rarely exceed the 5-digits. Besides, YouTube/Google is a foreign entity, which means that unless they want it, the German government has no access to them.

Another reason for the expected non-reaction is that dubbing has always been used as a way of censoring foreign information. In the movie "A fish called Wanda" for example the protagonist sings the first strophe of the German national anthem. In the dubbed German version this passage in the movie has been altered, because the first strophe of the anthem has been „phased out“ after World War II due to its overly nationalistic text (together with the second one which is about the out-dated German geography). This happened although the anthem does not specifically contain extremist verses.

The altering of the movie scene is just one example of censorship by translation of many - which by the way occured in former Western Germany, a place that used to be by far more relaxed in regard of the Overton window than the Germany of today is.

Nonetheless I see a certain chance for YouTube/Google to be punished in a severe way by the US legal system. YouTubes approach may be covered by the 1st amendment, but the translation service is a product. And while it may be usuable for free, this does not mean they have the right to cause damage to their customers or intentionally manipulate them politically.

As implicated in the title image, my comparison would be to the tobacco industry, a once thriving and powerful multi-billion Dollar industry - until it could be proven that they knowingly misled their customers and sold them poison.

False translations may not be poison for the body, but obviously for the mind. In my analysis above I prove that YouTube/Google is knowingly manipulating its tranlsation tool. All it needs to take them down is somebody who is being harmed by these manipulated German translations.

Perhaps someone like Tucker Carlson who is framed as an extremist, or perhaps someone like Tim Pool, whose work is distorted by the false translation of his commentary. Or maybe someone from Germany who is also a US-citizen but needs German subtitles when informing himself about political aspects in the US. I am sure, somebody is out there who fits the necessary profile for a court case.

Bottom line is, this is huge and I believe it could even be big enough to turn the Google founders into millionaires.