Gender bias in Google Translate.
Artificial intelligence is known for reinforcing stereotypes, this happens when you train a model on a large amount of data and then figure out that your underlying data has human bias.
With translation software, this bias is shown when a sentence requires the software to guess a gender when translating from a non-gendered language to a gendered one. Others have already written an entire paper on everything that goes wrong when you ask Google translate to translate from a language without gendered pronounce to one with gendered pronouns.
Unfortunately, I don’t speak any of those languages in this article I will focus on a Danish-to-English translation (since it’s the two languages I speak).
The word I will translate is kæreste. The most correct translation of kæreste would be my dearest or my beloved, however, in practice, the word means either boyfriend or girlfriend depending on the gender of your significant other.
So Google has some problems deciding if the kæreste is a boy or a girl and thus falls back on gender stereotypes whenever it is in doubt. So what can we do about this?
What about Bing?
So Google is not the only one who has a translation service, there is always the competitor Bing and surprisingly Bing has solved this problem by always translating kæreste to Boyfriend, while this makes it hard to accuse Bing of being biased there is a certain absurdity to this approach:
The problem with guessing.
Machine learning is when boiled down to its fundamental just using a huge amount of data to make the best guess, this means that while the bias is obvious it is also often the most accurate representation of the world. For example, I expected Google to hold the bias women are bad at driving but what I found was a far more accurate representation of reality, namely that women are more likely to have a small accident, while men are more likely to have serious accidents.
The same can be said for all the other examples, if you had to bet on the gender of the CEO of a random Danish company guessing on a man would make the odds in your favor, and if you have to guess who makes dinner tonight at a random Danish family guessing on the women would be the best bet. So Google is just making the best guess given the current state of reality.
Is this a problem?
I can follow the feminist argument that having Google maintains those gender bias is also reinforcing them and thus may prevent society from progressing, however, I can also follow the argument that AI should represent the world as accurately as possible and not how some data-scientist who spend to much time at a university wants it to be.
I don’t know how Woke we want AI to be, but I do know that the new Medium algorithm values comment engagement, so if you have a strong opinion on this, go ahead and write it.
However, in this particular case, there is a solution that should satisfy all parties.
The power of admitting that you may be wrong
Google already had a solution when people pointed out the problem with pronouns. They trained a new AI to detect when pronouns were used and then gave two possible translations.
The problem with this approach however is that there are thousands of ways new bias may sneak in. The reason I focus on the word kæreste is that it has nothing the do with pronouns, and the only way Google would know is if a Danish data scientist discover the problem. With all the possible combinations of different languages, they can not realistically map enough examples to make a reliable classification. The problem as I see it may also go way deeper than just gender bias.
As I see it the problem is that AI is making a guess when it should admit that it can not know the answer.
When writing this article I often saw the word kæreste change while writing the sentence as just small words could change the probability of boy or girl, the transition always appeared confident even though it only has 50.1% certainty, to illustrate here is a final example:
In case you didn’t notice the only difference between those two sentences is a comma, while it is impossible to know for sure what's goes on inside the black box of machine learning I guess that Google has learned that women are slightly better at using gramma than men and thus the sentence with the slightly more correct grammar must be women writing about her male partner. And while I do give Google credit for learning an AI with this advanced way of reasoning, I would say that sentence like this should call for Google Translate to admit that there may be two possible ways to translate this sentence and Google does not know which one is correct. Thus Google should solve this problem by giving the two most likely translations in all cases where ambiguity is present instead of relying on a bias that may be incorrect.
