In this blog post, we explore whether artificial intelligence can think like humans and make rational judgments beyond simple calculations, using various examples.
Deep Blue conquering chess, AlphaGo defeating Lee Sedol at Go, and the Haken and Apell algorithm proving the Four Color Theorem—a long-standing mathematical challenge—are just some examples of AI showcasing its limitless potential, astonishing us. These advancements are turning what was once considered possible only in past science fiction films into reality. As AI tackles increasingly complex and sophisticated problems, it is transforming not only the challenge to human intellectual capacity but also the structure of our society and our daily lives. Observing this, it appears AI is conquering, one by one, fields once considered the domain of ‘reason’ – the ability to solve problems using mathematical and logical tools. Some argue that AI already possesses reason surpassing humans and could pose a threat to humanity in the future. This concern is a recurring theme in films and literature, warning of the possibility that AI could replace humans or infringe upon humanity.
In his book The Most Human Human, Brian Christian argues that because AI is encroaching on the domain of human reason with its superior processing power, we must now pioneer the realm of emotion. He contends that as long as AI cannot mimic emotions—long considered uniquely human—emotion will be the final domain differentiating humans. However, this reaction raises a question: Can we definitively conclude that AI has encroached upon human reason merely by observing its victories in chess and its ability to handle complex mathematical calculations with immense computational power? Furthermore, human neurons possess an extremely complex structure, fundamentally different from the computational architecture of AI. It is said that this complexity makes it difficult to perfectly replicate human neural networks using simple structures like AI perceptrons. So, if AI’s processing method differs from that of the human brain, can we truly say they possess the same ‘reason’?
First, let’s consider chess, which Brian Christian referred to as a ‘human’ domain. While attributed to human error, it remains an undeniable fact that computers defeated humans in chess. However, this victory stemmed solely from rapid computation and statistical processing that evaluated countless possibilities – it was never a conclusion reached through the computer’s own thought. When the chess grandmaster and Deep Blue played, a widely recognized standard opening emerged in the chess world. The computer simply employed the most well-known solution for that opening, sacrificing its own knight to gain the upper hand. Observers joked that the commentators, seeing the computer use this solution without understanding its meaning, might only realize after the computer had finished its moves and ask, “Who took my knight?” Seeing this, I think it’s a matter worth pondering whether computers truly possess the ‘reason’ we speak of.
In chess, Brian Christian even mentioned the ‘book’ of chess, referring to its standardized patterns, because the possible starting positions and moves are relatively limited. Therefore, artificial intelligence could achieve good results based solely on ‘computational power,’ not ‘reasoning.’ So, what about Go, which has a vastly different dimension with an enormous number of possible ‘cases’? Experts predicted that Go, with its impossibly vast number of possible moves, would be too challenging for AI to win against humans for at least a decade, as simply having superior computational power alone would be insufficient. This is why AlphaGo’s victory came as such a shock to many. So, should we acknowledge that an AI that defeated humans at Go possesses reason?
According to Google, AlphaGo uses an algorithm called Monte Carlo Tree Search (MCTS). Instead of trying every possible move, this algorithm creates a database called a search tree through practice simulations, thereby reducing the number of ‘possible moves’ to calculate. After forming an extensive search tree through countless simulations, it selects the most valuable data from the relevant branches of the search tree during actual play to make its moves. This method is effective because while there are countless moves that do not violate Go’s rules, the number of moves that people actually consider meaningful and use in real games is relatively limited. The fact that AlphaGo reduced possibilities through the MTCT method was confirmed when Lee Sedol played moves not typically used by humans—moves not recorded in the search tree—causing the system to make errors. Thus, the MTCT method merely reduces computational load using a search tree; fundamentally, it remains a simple computational algorithm not vastly different from chess algorithms, as it still involves substituting numerous possibilities within a given framework. While it achieved unparalleled progress compared to existing chess algorithms by employing strategies to reduce computational load, it still largely remained within the methodology of substituting possibilities that conform to the rules. This methodology of substituting rule-compliant possibilities can demonstrate excellent capability within the given rules but carries the fatal limitation of being unusable beyond the scope of those rules. Ultimately, even AlphaGo, a cutting-edge AI, used this methodology. Consequently, its operational scope is confined within the rules, making it akin to a frog in a well confined by its rules.
Following its successes in chess and Go, AI has been applied to various games and now exhibits more evolved capabilities. For instance, in ‘2048’, a game requiring considerable thought, AI demonstrates superior performance compared to humans. ‘2048’ is a game developed in 2014, before the AlphaGo vs. Lee Sedol match. Each turn, players can move blocks in one of four directions. When blocks with the same number merge, they form a new block with the combined number. The goal is to create a 2048 block. Among the algorithms that performed exceptionally well in ‘2048’, the genetic algorithm stands out as a prime example. The genetic algorithm operates similarly to biological evolution: it slightly mutates numerous algorithms and pits them against each other. Algorithms achieving higher scores survive, while those failing to do so are eliminated. Initially, the AI moves randomly among the four possible directions. Later, it demonstrates proficiency by employing strategies like ‘pushing large numbers into corners,’ one of the high-scoring tactics. However, this merely represents random algorithms competing until the optimal one prevails; it absolutely cannot be said that the AI played the game with thought. Even in the case of ‘forcing the big number into the corner,’ while humans employ it through experience and intuition, the AI simply tried countless possibilities and happened upon the optimal algorithm. Thus, while genetic algorithms are strategic methods for exploring limited problem spaces, they remain confined to the methodology of substituting cases that fit given rules, much like chess or Go algorithms.
Observing AI play games, it may seem to possess human-like reasoning, but in reality, it merely statistically determines the method yielding the highest score by substituting all possible cases that fit the rules within a limited framework. In other words, while this methodology excels at ‘trying every possible approach within the framework of the rules,’ it cannot perform when the methodology exceeds the scope of the rules. Ultimately, viewing artificial intelligence as having caught up to human rationality may be an overinterpretation.
