Did Facebook Just Shut Down a Computer Because it got Too Smart??

Why didn't you "like" my holiday pics Dave?

Better Late Than Inaccurate
    I don’t often write about current affairs in Science for two reasons.  The first is that when a "news-worthy" Science story breaks, it gets splashed everywhere in the media so there’s no need for me to report it too.  The other reason is that I like to take my time with things.  When you hear a Scientific claim, the best thing to do is check it carefully, do some research, find original sources, learn the background etc.  Unfortunately, the media machine moves very fast so by the time I know what’s actually happened I’m usually behind the curve.
    And to be honest I like it that way.  I’d much rather be cautious when I hear a news story than comment within the hour.  Particularly if it’s complicated.  So, despite many people trying to persuade me to write more up-to-date stuff, I’m going to be stubborn.  Personally I value accuracy over expediency.
    One of the exceptions to these rules is when the "hype" over a story has gotten out of proportion, or that people are misunderstanding what actually happened.  In that case, I do feel more of an urge to try and put my thoughts out there, and this story of Artificial Intelligence (AI) gone haywire is a prime example.   
     You might have come across it a few days ago (31st July was when the story broke).  I ran across it on scaremongering Instagram feeds and ignored it, perhaps foolishly.  When it kept coming back I decided to look into it and see what the truth was.  It’s taken me a few days to get to grips as I’m not an expert on AI technology, but I’m pretty confident I can report with reasonable insight. 

Hello, would you like me to traslate this blog into Bocce for you? I am well versed in over six million different forms of communication and...

What Got Reported/Is Being Reported
     According to the headlines, facebook was doing research on AIs and successfully created two robots which possessed the ability to communicate with each other.  The robots struck up a conversation, but very quickly decided to abandon English and invented their own language which the programmers could no longer decrypt.
     The robots conversed in their secret “robot-ese” with increasing speed, hiding their conversation from us, learning as they went.  Panicked and frightened, the facebook programmers immediately shut down the software before it got too smart.  This is apparently the first instance of computers creating their own secret code system and attempting to outwit their human creators.

Genuine photograph of spooky computer ghost-monster-demons

What Actually Happened...
    Facebook, like many other companies who develop computer software, spend a lot of time researching chatbots.  Chatbots are programs designed to mimic human speech, useful for all sorts of things like voice recognition software or operating systems that talk back etc. 
    The way they work is by picking up on certain words, applying the basic rules of grammar and syntax, interpreting the message and outputting a logical response.  There’s a debate around whether this consistutes “speaking” a language, but a lot of chatbot software can be quite sophisticated.
    And chatbots aren't anything new.  In fact, there’s an annual competition called the Loebner Prize which has been running for twenty one years in which chatbots compete to try and convince a panel of judges they are human.  These tests (where someone is talking to a screen and isn’t sure if it’s a person or a robot) are called “Turing tests” and there are lots of chatbots which have reasonable success-rates at passing them.  Specific and detailed conversations are still impossible, but simple chats about the weather etc. can be simulated easily. 
    One of the things programmers of chatbots particularly like to do in order to road-test them, is to put two chatbots into conversation with/against each other.  Depending on your perspective this is either ingenius or hilarious.  The result is that the two chatbots communicate and try to understand each other’s usage of a language. 
    Obviously, when two chatbots talk they can end up exchanging complete gibberish because they don’t really understand English (that’s the whole point of the research, to see how close a simulation can get).  And that’s what these two chatbots ended up doing, the only difference was that their gibberish had a vague structure to it.  The language they were using was still English, just a slightly distorted version, which made little sense to any human reader.
     If you’re curious what their conversation looked like, here’s a short section of it.  In the spirit of the AI takeover, you should probably listen to the theme music from Terminator 2: Judgement Day (which I've helpfully included below) while you read the extract:

Extract from the Chatbot Conversation:

Bob: you i everything else . . . . . . . . . . . . . .

Alice: balls have a ball to me to me to me to me to me to me to me

Bob: i i can i i i everything else . . . . . . . . . . . . . .

Alice: balls have a ball to me to me to me to me to me to me to me

Bob: i . . . . . . . . . . . . . . . . . . .

Alice: balls have zero to me to me to me to me to me to me to me to me to

Bob: you i i i i i everything else . . . . . . . . . . . . . .

Alice: balls have 0 to me to me to me to me to me to me to me to me to

Bob: you i i i everything else . . . . . . . . . . . . . .

Alice: balls have zero to me to me to me to me to me to me to me to me to

Quaking in your boots right?
    Not so much "The Terminator" as "Jonny-Five suffering a tourettes outburst".  The chatbots started producing conversations like that and as a result, the programmers switched them off.  Not because they were being intelligent, but because they were being stupid.  And here's the thing: this sort of thing happens all the time.  Chatbots often start talking in garbled forms - the only reason this made the headlines was because it was facebook doing it and it's front page news when Mark Zuckerberg blows his nose. 
    The reason the malfunction occurred was because when the programmers wrote the chatbot they forgot to specify that the language had to stick to certain grammatical rules.  If we agree that a language has to have certain properties e.g. finite words, infinite sentences, recursivity, generativity, then there are 24 grammatical possibilities a language can take (linguistic logicians like Frederick NewMeyer have actually worked this out).  Of the 6,000 languages on Earth, only 15 of the possible grammar structures are actually used, with most languages sticking to one of 4. 
    In other words, almost every human language on earth conforms to one of 4 types, but there are 20 largely unused ones out there.  It’s no surprise that a mathematically-minded chatbot might select one of the others that might be far more efficient than ours.  Really, it’s no surprise that a computer would butcher our language...our language is a mess to begin with.
    So technically, the chatbots did start using their own language but it wasn’t an invented one.  It was just one of the other possible ones and they were still using English words.  There was nothing sinister going on and pulling the plug was done because the bots were failing to simulate our language...badly.  It wasn't so much a case of "Oh God the robots are sentient, quick pull the plug!" it was more "Oh damnit...Steve, the stupid bots are talking like idiots again, can you hit the stop button,  I can't reach it over my coffee!"  
    So no, we don't have anything to fear from facebook AIs getting too smart.  The worst we could say is that a robot who spoke like Alice or Bob would be extremely irritating. 

This blog was brought to you by Skynet. 

Now return to your homes and ignore all reports of a robot uprising.

Image credits:
HAL-9000: wallpapercave
C-3PO: theoldrobots
Scary computer: kkomando
Terminator 2 Sountrack composed by Brad Fiedel
Dr Evil: quickmeme
Skynet: pinimg