The Most Human Human

I seem to have a thing going for books that describe how the author tries to achieve some weird self-set goal (reading the Britannica in order to win a game show, reading through the complete OED, becoming US memory champion). In this instance the goal is to win the 'most human human' award that is given out at an annual Turing test.

To bring everyone up to speed: the Turing test was meant as a measure how well artificial intelligences perform. Judges have to have a 1:1 chat with a computer and a human confederate without knowing which 'intelligence' is only based in silico and have to make a call of which participant was human and which was the machine. If a computer can fool ~30% of the judges into believing that they are the human participant the program is said to have passed the Turing test and for the Loebner Prize there's the 'Most Human Computer'-Award for the program hat fools most judges, even if not passing the threshold. But as the competition is a zero-sum game (for each time the computer fools a judge into believing that it's human a human must be flagged as non-human) there's also the award for the human confederate that can convince the most judges that they are in fact human.

So that's what Christian is trying to win. And instead of just acting like a regular human being he sets out to prepare for the test by reading up on AI, computational linguistics and information theory to find the weak spots of current chatbots. Which is fun to read and one quickly notices that he has a professional background in computer science, philosophy and poetry. I really enjoyed reading about the different strategies applied to fake being human, like turning the chat into a shouting match, as insults are no longer dependent on a larger context but from a computer science perspective are more of a Markov chain, where your retort is only dependent on the last observed state.

Another fun topic was the application of the Shannon Entropy to text input on mobile devices via T9 and the different auto-correction algorithms around. Christian argues that these methods are shaping how you write, as they perform best on the vocabulary they were designed on and you thus unconsciously start to adapt to that vocabulary. Anecdotally this is just what I observe while typing away on my phone: Instead of the software adapting to my needs and figuring out what I wanted to say, I start to adapt to the idiosyncrasies of the auto-completion feature and try to predict what words the software can make most sense of, in order to avoid the software turning my intended writing into gibberish (especially for those üseless umlauts while typing in German. And I still have fond memories of T9 turning 'Sure, I'll come over and bring beer' into 'Sure, I'll come over and bring AIDS').

To borrow from Philipp's closings: Recommended for anyone interested in AI and computational linguistics (probably less if you're an expert in one of those fields).