Apple's AI has a porn problem

[big_D]

Aktuell | Maschinelle Übersetzung

Amusing article, with serious consequences.

Apple’s translation tool on macOS and iOS uses AI to work out what to translate. Unfortunately, it isn’t as advanced as some other translators and cannot recognise the source language, therefore you have to select the correct source.

One reader forgot to change the source language, when he was translating and copied an English sentence into the translator, when the AI thought the text was German. The result?

The original sentence was: “Mirror egg is very yummy in the tummy”,
Apple’s AI translation to English was: “The brunette is getting penetrated in the kitchen”

The c’t team looked into the problem and found some other example of text that get “pornified” when going through the translator:

Original: Happy dog is very young 3003
Translation: sexy brunette in black black black is happy to be banged

It looks like the AI has a thing for brunettes and has been watching too much Pornhub whilst waiting for translations. The c’t team reported the problem to Apple and Apple fixed the problem for the example sentences that the team provided, but according to the article, there are other sentences that get similar results.

This is amusing, but it also shows up the problems with AI and translation software in general. Translation software is great for simple sentences, but anything complex and they can quickly come a cropper.

I was writing documentation and my boss gave me a 4 hour deadline to translate a 20 page manual into German! My first thought was to run it through Google Translate and then tidy up the results and correct the small errors. After I stopped rolling on the floor laughing, I told my boss he would just have to wait for the translation, until I could get it done.

The problems? The biggest gaffes were:

Source: Do not open the case, high voltage inside
Translation: Gehäuse öffnen, Startstrom drinnen (Open the case, high voltage inside)

Source: Do not open the case, no user serviceable parts inside
Translation Gehäuse öffnen, nicht drinnen. (Open the cae, nothing inside)

It went on, about 20% of sentences were roughly accurate and could be used as a basis for a translation. I updated Google Translate with the worst offenders and went on to translate it manually.

Those translations, whilst absolutely hilarious, would have had dire consequences, if they had made it into the finished document.

Interestingly, I did try changing the English source sentences, to “don’t open the case” and it translated that accurately. It looks like Google Translate can cope with abbreviations properly, but it has problems with formal English.

[PHolder]

I’m sure, if Apple lets Siri unionize, and raises her wages, she’ll get it together more quickly

On the other hand, there might be a market opportunity here for a trashy romance company

[knewman]

Pretty funny translations. Hopefully the fixes didn’t burn too much of your day!

Have to say I agree with the dire consequences though. Not AI specifically, but software development in general. I fear it’s getting to some sort of convergence where modern software programming is far too complex for humans, yet we’re putting greater weight on the backs of dubiously developed software systems. I recently had a scary moment with the collision mitigation software in my car in which the car nearly caused a collision. The manufacturer is busy chasing what they believe to be a faulty sensor, while I’m confident this was a software flaw in their code.

It’s all fun and games for a translation program, but when these development practices are adopted in systems that can put people in harms way, the fun stops.

[PHolder]

The problem with machine learning, as it is currently practiced, is that we don’t really understand how the “machine” actually works to come to a conclusion. (This is akin to the fact that we don’t really understand how the human brain works either, and the ML based on neurons is kind of replicating a human, to some extent.) I think there are a number of possibilities.

One possibility is that we put learned machines (those that have been trained via machine learning) in a box where they’re only allowed to interact with problems we have previously tested them with. This doesn’t seem as useful in everyday life, but it seems hella safer.

Another possibility is we restrict our reliance and trust in a ML device for many months or years until we know it has succeeded in more and more new scenarios. In essence trust but verify where the amount of verification falls off over time.

Another possibility is that we strictly separate the training of multiple machine learning devices, and we group them in such a way that they have to come to consensus before they’re allowed to provide the result back. This may not be foolproof, but it should act like decision by committee. The hope is the “experiences” that led to the training of the ML device are varied enough between the devices that they all have different biases that in essence cancel each other out.