Before we go deeper into the discussion of the philosophical issues that are created by Artificial Intelligence systems (see here for a list of posts), it will be useful to get a rough idea of the kinds of systems we are talking about. This is not an exhaustive list by any means, and it does not even represent every kind of AI system. It is a selection of systems that, in my opinion, pose interesting philosophical questions. We will use some of these systems as examples in later discussions.
Deep Blue, IBM Watson, and AlphaGo
It’s only twenty years ago, but it’s another universe. In 1997, Bill Clinton was president of the US, Nokia was the largest mobile phone manufacturer in the world, and the first iPhone was ten years away in the future. People still talked of the greenhouse effect (instead of global warming), and you’d probably read this through your dialup modem on a Pentium II PC with 64 MB of main memory. You’d get your Internet access through AOL (like half of all American households). Oh, and forget about googling anything, because Google would be founded only a year later.
Still, already at this point computers would start winning against humans in games where humans were considered invincible. In May 1997, the IBM computer Deep Blue won the rematch against the then strongest chess player in the world, Gary Kasparov. It was a narrow win, and up to the last game the two opponents both had 2.5 points. On May 11, Deep Blue won the sixth game of the rematch, and became officially the first computer to win in chess against a human world champion under official tournament conditions and time controls. It was debated for a while whether there was any cheating involved from the side of IBM, or whether Kasparov lost because he didn’t play well, rather than because the computer had become invincible. But whatever the reason was for Kasparov’s loss in 1997, since then computers have clearly become powerful enough to easily defeat any human opponent in chess, and nobody would dispute this today.
IBM’s computers earned a much more complex win in 2011, when IBM’s computer Watson won against two human players in Jeopardy. The game, for those who don’t know it, presents the player with questions like: “Sakura cheese from Hokkaido is a soft cheese flavored with leaves from this fruit tree.” The player then has to guess the missing word, in this case the name of the fruit tree (and present the answer in form of a question, but this is a purely formal requirement that does not affect the difficulty of the game). The questions were presented to Watson in text form (so the machine didn’t have to recognise spoken text, which would have been harder in 2011 than it is now). Still, as the sample above shows, the questions require quite a bit of the ability to understand natural language, to work out what is actually asked, and then to look up the answer in a very extensive database of facts. Watson won against the two human champions with a very big lead, and has since been developed as a knowledge processing machine for other, more serious applications. ‘Watson’ is still used by IBM as a label for machines that today can recognise objects in pictures, perform medical diagnoses, and communicate with customers of various companies in natural language over the phone.
The third important milestone in game AI development (and many say, the last) came in 2016 and 2017, when DeepMind’s AlphaGo easily and decisively won nearly all games against two top human players of the ancient Chinese game of Go. In March 2016, AlphaGo won 4 out of 5 games against a top human player, Lee Sedol. One year later, the program, now improved with more playing experience, won all three games against the strongest human player, Ke Jie (Future of Go Summit, 23-27 May 2017). At the same event, there was also a Team Go game, where five human players together played against AlphaGo: Chen Yaoye, Zhou Ruiyang, Mi Yuting, Shi Yue, and Tang Weixing. They lost together against the program.
Computer winners, human losers
Looking at this history, one thing that becomes obvious is that the clear and final defeat of humans against AI players does not seem to deter humans from playing these games. One would perhaps expect that chess championships would lose some of their appeal after the defeat of Kasparov in 1997. More generally, one might expect that games that humans can’t possibly win against a computer might in the long term become less attractive to human players. What’s the point of playing a game professionally, one might say, if even the strongest human player is not the strongest player any more? If there’ll always be someone far stronger than even the best human champion?
But, perhaps surprisingly, this hasn’t happened. Even after 1997, we have chess championships, and the current chess world champion was just seven years old when Kasparov lost to Deep Blue. When he decided to pursue this career, it was already clear that there was no possible winning against computers. Still, this did not keep him from becoming a professional chess player, and neither did it deter thousands of others. So what happened here? Why are humans not as upset by losing to computers as one might expect? Why do they play on, even after computers have “solved” a game?
One way to look at the question is by comparison with other sports. It was always the case that humans ran slower than lions. They also ran slower than humans on horseback. They also run slower than cars. Still, this has not stopped human runners from pursuing marathons, or attending the 100m races in the Olympic games. We don’t usually question the point of attending the Olympics on the argument that cars can run faster.
This provides a clue. Human sports, despite appearances, are not focussing on the outcome only, in the sense that some numerical measure of performance is the only thing that counts for the athlete. If it were so, then perhaps all sports would have been given up when people realised that other entities (animals, cars, computers) perform better than them regarding the particular measure. But if athletes don’t focus on the outcome, what is it then that they pursue when they take up a sport?
What is the true purpose of a game, or a competitive sport, for those who participate in it?
Why do humans play games?
Here are three obvious answers to the question why we, as humans, play games:
- To enjoy the game, to be entertained.
- To compete against others (a social function).
- To challenge oneself to improve.
These are three distinct reasons. I might play chess or Go because I am bored and I want to spent an hour doing something that’s interesting to me and that I find entertaining (reason 1). Or I might want to compete against others in a match, to see who is stronger, and to feel the thrill of winning against a strong opponent (reason 2). This is different from the first reason. If I play for entertainment against my little daughter (reason 1), I might not insist on winning. In fact, I might try to lose in order to motivate my daughter to play against me again, and to make sure that she doesn’t feel bad about the game. Reason 2, instead, focuses on the built-in sense of competition that primates (and humans) have to such a high extent: wanting to be the strongest, fastest, cleverest, and generally best admired among all others one competes against, and thus to rise to the top of one’s social hierarchy. This is not a pretty trait, but human societies are built on it, and it’s hard to deny that competition motivates much of our everyday behaviours: from ranking in schools, to promotions at the workplace, to the way we display status and wealth through fashion, the cars we drive, and the houses we live in. Competitive games are just another way to earn and display rank.
And finally, I might be playing in order to challenge myself to improve at a skill that I find admirable, and that I lack to some extent. We often learn things for that reason. People take drawing classes, learn a foreign language, or learn to swim, to garden, or to play chess for the reason that this gives them a measurable way to improve, to feel that they are getting better at something, and to experience the pleasure that comes from that feeling of gaining competence in some new area of human activity. Improving at something is in itself already a rewarding experience, regardless of any other, external gains that may or may not come from it.
Are computers actually playing?
Realising why we, as humans, play, we could even use this as a definition of a game: We might say that a game is an activity that is performed voluntarily by individuals in order that they be entertained, or in order to compete with others and gain rank, or in order to challenge themselves to improve in performing this activity. In all three cases, positive emotions are derived from achieving each goal through playing the particular game.
Now we can see two things:
First, the human goals of game playing are not endangered by computers playing these games. Clearly, I can enjoy a game, I can be entertained by it, even if the game is played (better) by computers. I can also enjoy a game that I know I will lose (for example, if I play against a chess program). If my goal is to pleasantly pass the time, I can still do so (either against a human or against a computer) even knowing that computers can, in principle, always beat humans in that game. This realisation simply does not affect my enjoyment of actually playing the game. Similarly, if my aim is to compete socially for rank, then I will not play against a computer (since computers don’t participate in human ranking schemes), and I will not measure my performance against any computer’s. Instead, I will compete with a human for a place on a human rank list. The fact that computers might be better at some activity does not affect my pursuit of rank amongst my peers. Finally, if my aim is to improve my own abilities, again it will be irrelevant to me whether computers can play the same game better than humans. I’m simply not interested in this fact, since my motivation is to become better by learning to play this game. A computer can help me do this, but it cannot deter me from doing it. So there is really no reason why humans should abandon competitive games just because computers play them better. (Same with running: the enjoyment, the competitive, and the self-improvement aspects of running are not affected by the fact that cars run faster!)
Second, we can question whether computers even can be said to “play a game” in the same sense as we do. If we define “playing a game” as I suggested above, then computers cannot even be said to “be playing” chess or Go. We use this kind of language as an abbreviation, but it is not really true:
- Computers don’t play for their own enjoyment or entertainment. Playing is, for them, an activity they are programmed to perform, not different from calculating a column in a spreadsheet, or driving a self-driving car. None of these activities cause any enjoyment or annoyance to the machine. They are all just activities performed for no other reason than that the machine was ordered to perform them.
- Competing against others (human or not) is also not a motivation a machine might entertain. Machines don’t even realise that they are competing, since they (at least for the moment) lack any awareness of their own identity. A game, to them, is just a series of complex calculations, and they have no sense of competition, no desire to win, and no consciousness of any social hierarchy that might be affected by them winning a game. Deep Blue didn’t suddenly “feel great” because it beat Kasparov.
- And finally, although (at least some) machines do improve by playing games (AlphaGo learned many of its Go skills by playing against itself), such improvement is not the goal of the machine itself. Rather, the programmer or operator orders the machine to play games. The machine itself is not motivated to play out of a wish for self-improvement.
Thus, we might argue that what machines do when they “play” chess is, in fact, a totally different activity from what humans do when they play chess. Machines don’t actually “play” anything, since they lack the motivational infrastructure for playing games. They perform a behaviour, but this, in itself, is not yet playing. In the same way, a car that drives along a marathon track is not, in any valid sense, “running” a marathon. It does something entirely different, namely driving along a marathon track. And a computer that performs the sequence of commands that implement legal chess moves is not actually “playing” chess. It is just performing legal chess moves. This is an important distinction that is almost always overlooked, since we are so fond of using phrases like “AlphaGo plays a game,” which are fundamentally misleading as to what actually happens when AlphaGo decides to perform a move on the Go board.
Let’s leave it at that for the moment! In the next few posts we will briefly discuss other areas of application for AI systems, and we will have a look at what philosophical problems they pose! Stay tuned.