Human Beats Artificial Intelligence in Go: Prof. Stuart J. Russell Highlights Fundamental Flaw in AI Like AlphaGo

In a stunning turn of events, a top-ranking AI system was soundly defeated by a human player in the strategic board game Go. This marks a significant reversal of the 2016 victory by the computer, which was widely celebrated as a landmark achievement in the field of artificial intelligence. The victorious player, Kellin Pelrine, holds a rank just below the top amateur level and was able to exploit a previously unidentified vulnerability in the AI system that had been detected by another computer. The games were played without any direct computer assistance, and Pelrine emerged as the clear winner by winning 14 out of 15 games.

Financial Times mentioned that this victory, which has not been previously disclosed, brought to light a flaw in the top Go computer programs that is common to many of today's prevalent AI systems, including San Francisco-based OpenAI's ChatGPT chatbot. The strategy used to win the Go match was proposed by a computer program that investigated the AI systems to discover any weaknesses. Pelrine, who executed the proposed plan with merciless precision, was able to use the tactics to put the human player back on top of the Go board.

AI System's 'Blind Spot'

According to Adam Gleave, the CEO of FAR AI, a research company based in California that created the program, they were able to easily exploit the system. The program played over a million games against KataGo, one of the leading Go-playing systems, to uncover a "blind spot" that could be exploited by a human player. Pelrine stated that the strategy for winning, which the program uncovered, is not overly complicated and could be learned by an intermediate-level player to defeat the machines. He successfully employed this approach to win against Leela Zero, another leading Go system.

Despite relying on tactics suggested by a computer, the resounding victory occurred after seven years of AI seeming to hold an insurmountable advantage over humans in what is widely regarded as the most intricate board game of all. In 2016, AlphaGo, a system developed by Google-owned research company DeepMind, beat the world Go champion, Lee Sedol, by a margin of four games to one. Sedol later attributed his retirement from Go three years after the event to the emergence of AI, describing it as an unbeatable entity.

Although AlphaGo is not accessible to the public, the systems that Pelrine overcame in his recent victory are deemed to be on a similar level. The game of Go is played on a 19x19 grid, with two players taking turns to place black and white stones on the board in an attempt to encircle their opponent's stones and occupy the most significant amount of space. The vast number of possible move combinations makes it impossible for a computer to evaluate all potential future moves accurately.

Uncertain Defeat of AI

Pelrine's winning tactics involved slowly and methodically constructing a large "loop" of stones to surround one of his opponent's groups while simultaneously distracting the AI with moves in other areas of the board. Even when the encirclement was nearly finished, the Go-playing bot was unable to detect its vulnerability, according to Pelrine. Pelrine remarked that a human player could quickly notice the weakness in the AI system.

The revelation of a vulnerability in some of the most sophisticated Go-playing machines suggests a fundamental flaw in the deep learning systems that are the foundation of today's most advanced AI, according to Stuart Russell, a computer science professor at the University of California, Berkeley. He stated that these systems can only "understand" specific scenarios that they have encountered before, and they are unable to generalize in the way that humans can. Russell also said that this development demonstrates once again that we have been too quick to attribute superhuman levels of intelligence to machines, as reported by Arstechnica.

The exact reason for the Go-playing systems' defeat is uncertain, as stated by the researchers. One possible explanation is that the method employed by Pelrine is not commonly used, and thus the AI systems had not been taught enough similar games to recognize their vulnerability, according to Gleave. He noted that it is typical to uncover flaws in AI systems when they are exposed to adversarial attacks, as in the case of the Go-playing computers. Despite this, Gleave stated that we are witnessing the large-scale deployment of AI systems with minimal verification and that it is crucial to take a more cautious approach.

Check out more news and information on Artificial Intelligence in Science Times.