We've written at length on this blog about the perils of relying on machine translation. Errors in the translation of restaurant menus or ads are humorous but almost entirely without consequence. However, in high-stakes situations such as a courtroom (or interactions with law enforcement or the emergency services), the consequences can be serious for both the parties involved and society as a whole.
While we're now pretty well accustomed to tools like Google Translate and the native translation systems built into Twitter and other apps, they are still enormous undertakings in terms of computer processing power. Indeed, training a translation app can have the same carbon footprint as a long-haul flight.
In technical terms, these tools are simply algorithms. This is a step up from literal word-for-word translations as the AI 'learns' which words and sentences occur frequently together and translates using probabilistic estimates. However, what these tools can't yet do is add meaning, content, tone and intention. So, to generate the translation, it simply takes an educated guess based on the inputs it has been trained on. This part of the process can generate unintended and unpredictable outputs.
Some examples from recent years illustrate the dangers of relying on these apps:
When social niceties go wrong
A young Chinese man used an app to suggest to his female co-worker's Korean husband that they should socialise soon. The app failed to take into account the social norms of both countries (where people are given titles based on age and status) and imported a term used colloquially to refer to a woman of low repute. This resulted in a fracas, and the Chinese man later sought vengeance and killed the Korean man.
Good morning, bad day
A young man working in construction on the West Bank took a photo of himself leaning against a construction vehicle. He captioned the image "good morning" in Arabic, however, Facebook (the network to which the image was posted) rendered this the verb " to hurt " in English and "to attack" in Hebrew.
Clearly, in a country with such a febrile atmosphere, this was treated seriously. The man was arrested and questioned. The man was released after Police realised the error.
Probable cause... of a trial's collapse
In Kansas, a police officer used Google Translate to ask a Spanish-speaking suspect for permission to search the car. Already on legally shaky ground, even if the translation was bulletproof, the case was later thrown out as the translation was inaccurate.
It's no surprise that these computer generated translations go wrong from time to time. Language is complex and translating even more so. Syntax, sentence length, jargon, grammar, tone, intonation and context are all vital.
While these models are improving all the time, and are undoubtedly useful in non-professional scenarios such as asking for directions when on holiday, there is, as we can see, room for error.