AI Pidgin
In linguistics, pidgins are language forms that arise when two or more groups have to communicate but do not share a common language. Importantly, pidgins are language forms, i.e. they are not languages per se and nobody speaks them natively. One can make the analogy to domain-specific languages (DSL) in the typology of programming languages: both DSL and pidgins fulfil a certain function and are well-suited for that function only. Interesting examples for pidgins: Russenorsk, Sabir.
What if we make a pidgin to communicate with our pseudo-AI assistants? I am not going to comment on the commercial availability of the idea, but I will daydream (or nightdream — it is darkish outside) about such a hypothetical language form:
1. Phonological inventory should be kept to a minimum. On one side, to make our pidgin as easily pronounceable as possible for as many people as possible. This would mean limiting ourselves to the three most common plosives /p/, /t/, /k/; the nasals /n/, /m/; and a couple of fricatives, e.g. /s/ and /x/. We could also add a generic liquid, which might be pronounced as /l/, /r/, /w/, or whatever else is found in that range. No voiced obstruents, aspiration must not be taken into account, no typologically rare classes such as clicks or dental fricatives. Of the vowels I would only keep the generic three /a/, /i/, /o/, while discarding distinction in roundness or nasality.
On the other side, a small inventory with no two phonemes too close to each other should make speech parsing easier for Pipi, our pseudo-AI assistant (just decided to give it a name).
2. Free word order. Human languages have all kinds of word order and I have experienced how hard is it to be forced to speak out the words in pre-specified patterns. Pipi, on the other side, does not care. One interesting consequence is that our pidgin cannot have direct objects: the role of verb patients would have to be expressed via a preposition, just like all the other roles, perhaps with the exception of the agent. Why exactly prepositions when we can choose from so many different types of adpositions and case markings? Because it is most common.
3. Morphological categories should be only a few: two numbers, a handful of tenses. And these must be optional, i.e. it should be find to omit the tense affix if you do not care when be the verb.
Sounds far-fetched? Chances are that you think that nobody is going to learn a new language, be it only a form, to issue commands to GoogleNow. But I would say that pidginisation is already happening. Like all other pidgins, this one is being derived from living language(s), and like many other pidgins, this one is being derived from English in particular. I would say that humans issuing commands to computers are in a similar position as Australian farmers issuing commands to Papuanesian contract workers. You only know your language and you want to say something to an intelligence coming from a mind-bogglingly different background. Naturally, both of you will end up agreeing on a small set of basic vocabulary, will strip down the grammar to the bones, and will try to pronounce words clearly while accepting the other party's phonological variations. Think about it, do you google for "What is the timetable for trains between Stuttgart and Tübingen?" instead of "stuttgart tubingen trains"? Furthermore, I predict that soon we will have newly coined words aiming at simplifying common commands.
You can argue that computers are becoming better and better at understanding natural human languages. Absolutely true; in fact, I believe that near-perfect parsing of human output might be here any year now. But it does not mean that people will use it for everyday interaction with their computers. I personally would not want to lead a polite conversation with my augmented search engine. At least not until it becomes a real AI.