AlphaGo – move 19, or the triumph of “dehumanised” thinking
The whole world followed the game of Go with genuine interest. It pitted world champion Lee Sedol against Google’s AI, AlphaGo. More than the final result, which did see the machine come out on top under the very surprised gaze of every AI specialist watching, one detail really caught our attention.
In the second game, AlphaGo plays its 19th move and places its stone in a spot no commentator can make sense of. Its opponent is thrown. So much so that he leaves the room to splash water on his face and gather himself. It’s a configuration he has never come across in all his years as a world champion. At the time, the commentators called the move strange, and even wondered whether it was a bug [1]. A completely unexpected move that, to a human, makes no sense at all.
AlphaGo runs on a neural network trained on millions of games played against itself, and on games it has watched. It picked this move out of a tree of possibilities, because its decision algorithm had worked out that it offered the best shot at winning, given the moves already played and how likely the moves to come were. In hindsight, everyone agrees that this “unique”, “creative” move laid the foundation for the machine’s win in that game.
The electronic copilot - intelligence you can’t make sense of
In our Human Factors work in aeronautics, one case has come up time and again [2]: the electronic copilot of the Rafale. The project ran around twenty years ago now. The idea was to give the crews of two-seater Rafales a way to eventually replace the second pilot, who was responsible for navigation in particular, with an electronic copilot powered by an artificial intelligence.
The methodology was a textbook example of HF work, with great care taken to capture the copilot’s activity faithfully, in all its variety. Many pilots got fully involved in handing over their know-how, the goal being to build the knowledge base of a standalone expert system. From what we heard, the expert system was, technically, perfectly functional.
https://youtu.be/CGEaOg0l_KA?t=59
Even so, the first trials with pilots turned out to be a real disappointment. The verdict was that this electronic copilot just wasn’t intelligible. Pool all those human intelligences together and you get a system whose intelligence is too artificial to be understood. Collaborative decision-making rests as much on individual human expertise (the kind that made up the expert system’s knowledge base) as on a set of non-technical skills: social cues, the stereotypes we carry around that give us a frame for anticipating behaviour. Everything that an artificial intelligence simply couldn’t carry across. In the end, the project was abandoned.
A robot woman behind the wheel…
Stereotypes are a major concern as autonomous vehicles edge closer to the road. As the article by Munduteguy & Darses (2002) shows, decisions on the road are heavily shaped by the stereotypes drivers hold about their fellow drivers. Stereotypes about women, older people, drivers of big-engined cars or learner drivers all feed into our read of the situation in front of us and, with it, how we adjust our own behaviour! And never mind if those stereotypes are completely wrong. We get ready based on how we expect the driver we’re dealing with to behave. That stereotype is one of the only ways we have to anticipate.
But this ability to stereotype other people’s behaviour falls away with autonomous vehicles, which have no driver. So it raises the question of whether we can understand the artificial intelligence that moves the vehicle, and therefore whether we can share the road with it in unusual situations. To find out more about our work on autonomous vehicles, see our study on driver experience and automation through a usage-centred approach.
Artificial intelligence and literature
Each of us has no doubt had a quiet chuckle at Google Translate’s rough, seemingly clumsily literal output. Yet the tool is far more sophisticated than the results often let on.
Far from a simple word-for-word swap, Google Translate works on a principle known as Statistical Machine Translation (STM). Just as the tool Linguee does, it analyses millions of texts translated into the source and target languages, then uses statistical models to pin down the context of use of semantic elements and put forward fitting translations. The approach sometimes produces a kind of poetry (opens in new window), but it doesn’t always produce meaning.
Conclusion on artificial intelligence
Artificial intelligence is built on data, the data of human behaviour. But the traces of human activity, even in context, can’t all be fed through automatic processing in a way that captures the behaviour that produced them. Quantitative approaches are bound to overlook the individual behavioural dimension. It’s too complex; variability won’t fold neatly into large numbers. That’s exactly why a Human Factors approach like the one we practise at Akiani, on small samples but out in the real world, surfaces insights that marketing research can’t always reach! Our methodology also draws on the principles of ecological design and eco-design to build more human, more sustainable solutions.
Another of our articles on Artificial intelligence is here.
[1] In 1997, IBM’s DeepBlue artificial intelligence beat Kasparov at chess thanks to a bug. On the 44th move of the first game, and “although it was programmed to solve up to 200 million positions per second, the machine proved unable to choose one. So it was entirely at random that it moved a pawn, leading to its sacrifice. Now having a completely different perception of Deep Blue, Kasparov totally changed his strategy.” “Looking for the trap, the champion thought that Deep Blue possessed a superior intelligence, which had the effect of unsettling him right to the end of the match, which he lost (1 win, 3 draws, 2 losses).” Source (opens in new window).
[2] We’re telling it as best our memory has kept it. If anyone who worked on the project at the time would like to correct or add anything, we’d be very glad to hear it!
