The AI Paradox: Making Speaking a Language More Complicated Than It Really Is

| 7 September 2015 | Sebastian Feller

Recently I reviewed The Interactive Stance by Jonathan Ginzburg which was published in 2012 under Oxford University Press. After having read the first couple of pages, I breathed a sigh of relief. The book felt like a great leap forward in the natural language understanding business. Finally, someone in the AI world dares to leave behind the orthodox system view of language and models language as what it actually is: communicative actions carried out by two or more human beings in social interaction.

Great, I thought, let's see what's next. With great enthusiasm I browsed through the next couple of pages but, admittedly, my mood began to darken. I read about things called records which are made up of fields which are made up of labels and value pairs. I read about record types which are distinct classes of records. I read about the Type Theory with Records, or short TTR, which serves as the logical base for a framework called KoS. KoS supposedly represents meaning in interaction. KoS feeds into another theoretical construct, the dialogue gameboard. The dialogue gameboard is a sort of computable representation of
the cognitive states of each dialog partner in a given conversation. The details of the gameboard are highly technical and require an expert understanding of formal semantics and higher-order logic for one to stay on top of what is going on there.

Whenever I read this kind of stuff, I am thinking to myself: is speaking a language really that complicated? And is making a machine speak a language even more complicated? The whole AI business circles around a paradox: the linguistic theories and representations of language used for the machine are tremendously elaborate; however, the performance of the machine is usually meager.

Here a recent example: Last year, Eugene Goostman, a chat robot that impersonates a 13 year old Ukrainian boy, supposedly succeeded in the Turing Test. Eugene tricked 33% of test chatters into believing that it was a real person. The University of Reading, which organized this Turing Test event, claimed a historic breakthrough in the AI world.

That claim, however, has been received as largely controversial. According to an article on the website of the German newspaper Zeit (accessed on 18 Aug 2015), neither the details of the exact test setting nor the transcripts of the conversations have been made public by the organizers. Doug Aamoth, who writes for Time magazine on tech-related topics, published the transcript of an interview with Eugene Goostman on Time.com (accessed on 18 Aug 2015). Reading through the conversation, I got a weird feeling right from the start. At almost every turn, Goostman changes the topic of the conversation. It asks questions without any clear intention. The chatbot is literally all over the place without ever taking an interest in any of the answers provided to it by the human counterpart. It's a bit like Oh, so how did your last job interview go? - Well, it wasn't really that great. - Yeah, right. Do you like ice-cream? I hope this is not how real people talk to each other. But who knows. Times change.

Another annoying thing about Goostman is its bad memory. After a short while, it starts repeating its own questions over and over again. It creates that Groundhog Day experience that Bill Murray's character had to go through in the movie. That is where believability drops to close to zero.

To be fair, I am not sure if the version of Goostman in Aamoth's interview is identical with the one used in the Turing test. Be that as it may, the performance in both cases does not seem to me as remarkably fantastic as it was claimed to be by the organizers. How can we translate a 33% result in the Turing test anyway? Does that mean that Eugene Goostman is a little bit human? But isn't that the same thing as with being a little bit pregnant? It just doesn't work that way, does it?

Coming back to the AI paradox and the question whether speaking a language is really that complicated, I throw you another paradox: yes and no. It gets complicated for a machine that tries it for all the wrong reasons like winning a computer science competition about tricking people into false beliefs; in contrast, it is fairly easy for a machine that has real intentions and goals.

Now, what do I mean by real intentions and goals? Before every word there is an intention, a goal that the speaker wants to reach. Intentions are the key to language; they are, in a way of speaking, the fuel on which the language engine runs. From this point of view, it is not so much about modeling language per se, if one wants to understand language. Rather, the larger part of language is understanding the goals people seek to achieve through using it. This is where I believe the AI business is going wrong. Instead of modeling intentions and goals, AI people are exclusively into language.

Thinking about goals means thinking about the future: a future that one desires. Apparently, Eugene Goostman does not think about the future all that much. Instead it reminisces about the past, namely the last communicative move of the human chat partner. And that is exactly where it fails. AI research should look into adequate and computable representations of a “desired future”. Only then will machines be able to use language in a human-like way.

So let me ask you my final question for today: am I, the author of this blog post, human or a machine? Your opinion is very interesting… Keep going. And I forgot to ask you where you are from…

Comments: fellers(a)ihpc.a-star.edu.sg

Tagsdialogue, artificial intelligence, Eugene Goostman, intention, future, technology


About Us

Mission and Strategy



Former Officers







FILLM Events

Member Events



Book Series

Congress Proceedings

FILLM Newsletter


Latest Posts





FILLM Committee