In 2016, a computer named AlphaGo made headlines for defeating then-world champion Lee Sedol at the ancient and popular strategy game Go. The “superhuman” artificial intelligence, developed by Google DeepMind, lost only one of five rounds to Sedol, drawing comparisons to Garry Kasparov’s 1997 chess loss to IBM’s Deep Blue. Come on, who involves players competing by moving black and white pieces called stones with the aim of occupying a territory on the game board, had been considered a more insoluble challenge to a machine adversary that chess.
AlphaGo’s victory sparked a lot of angst about the threat of AI to human ingenuity and livelihoods, much like what’s happening now with ChatGPT and its kin. In a 2016 press conference after the loss, however, a moderate Sedol offered commentary with a core of positivity. “His style was different, and it was such an unusual experience that it took me time to adjust,” he said. “AlphaGo made me realize that I need to study Go more.”
At the time, European go champion Fan Hui, who had also lost a five-match private round to AlphaGo months earlier, said Wired that the matches made him see the game “completely differently”. It improved his game so much that his world rankings “skyrocketed”, according to Wired.
Formally tracking the messy process of human decision-making can be difficult. But a decades-long record of professional Go player movements has given researchers a way to gauge the human strategic response to an AI provocation. A new study now confirms that Fan Hui’s improvements after completing the AlphaGo challenge weren’t just a fluke. In 2017, after that humble AI victory in 2016, human Go players gained access to data detailing the moves made by the AI system and, in a very human way, developed new strategies that led to higher quality decisions in their game. Confirmation of the changes in the human game appears in the findings published March 13 in the Proceedings of the National Academy of Sciences of the United States.
“It’s amazing how human gamers have adapted so quickly to incorporate these new findings into their own game,” says David Silver, Principal Investigator at DeepMind and AlphaGo project manager, who did not participate in the new study. “These results suggest that humans will adapt and build on these findings to massively increase their potential.”
To determine whether the advent of superhuman AI has caused humans to generate new game strategies, Minkyu Shin, an assistant professor in the Department of Marketing at City University of Hong Kong, and his colleagues used a database of 5.8 million moves recorded in games from 1950 to 2021. This record, preserved on the website download go games, reflects every move of go games played in tournaments as far back as the 19th century. Researchers began analyzing games from 1950, as that was the year the modern Go rules were established.
In order to start going through the massive record of 5.8 million game moves, the team first created a way to gauge the quality of decision-making for each move. To develop this index, the researchers used another AI system, KataGo, to compare the success rates of each human decision to those of AI decisions. This extensive analysis involved simulating 10,000 ways the game might play out after each of 5.8 million human decisions.
With a quality rating for each of the ongoing human decisions, the researchers then developed a way to determine exactly when a human decision during a game was novel, meaning it had not previously been recorded in history. of the game. Chess players have long used a similar approach to determining when a new strategy in the game emerges.
In the Go game novelty analysis, researchers mapped up to 60 moves for each game and marked when a new move was introduced. If it emerged, say, on move 9 in one game but not before move 15 in another, then the first game would have a higher novelty index than the second. Shin and his colleagues found that after 2017, most of the moves the team defined as romance happened at the 35th move.
The researchers then examined whether the timing of new moves in the game was followed by increased quality of decisions – whether such moves actually improved a player’s advantage on the board and the likelihood of a win. They especially wanted to see what, if anything, happened to the quality of decisions after AlphaGo beat its human challenger Sedol in 2016 and another round of human challengers in 2017.
The team found that before AI beat human champions in Go, the level of human decision quality remained fairly even for 66 years. After that fateful period of 2016-2017, decision quality scores started to climb. Humans made better game choices – maybe not enough to consistently beat superhuman AIs, but still better.
Novelty scores also increased after 2016-2017 due to humans introducing new moves to games earlier in the play sequence. And in their assessment of the connection between new moves and higher quality decisions, Shin and his colleagues found that before AlphaGo was successful against human players, new human moves contributed less to good quality decisions, on average, than non-novel moves. After these landmark AI victories, new moves introduced by humans in games contributed more on average than previously known moves to better decision quality scores.
One possible explanation for these improvements is that humans were memorizing new move-set sequences. In the study, Shin and his colleagues also assessed the extent to which memorization could explain the quality of decisions. The researchers found that memorization would not fully explain improvements in decision quality and was “unlikely” to underlie the increased novelty observed after 2016-2017.
Murat Kantarcioglu, professor of computer science at the University of Texas at Dallas, says these findings, taken with work he and others have done, shows that “clearly, AI can help improve human decision-making.” Kantarcioglu, who was not involved in the current study, says the AI’s ability to process “large search spaces”, such as all possible moves in a complex game such as Go, means that the AI can “find new solutions and approaches to problems”. .” For example, an AI that flags medical imaging as suggestive of cancer could cause a clinician to take a closer look than before. “This in turn will make the person a better doctor and prevent such mistakes in the future,” he says.
One snag – as the world is seeing right now with ChatGPT – is the issue of making the AI more reliable, adds Kantarcioglu. “I think that’s the main challenge,” he says.
In this new phase of concerns about ChatGPT and other AIs, the findings offer “a hopeful perspective” on AI’s potential to be an ally rather than a “potential enemy in our journey to progress and improvement,” Shin and his co-authors wrote in an email to American Scientist.
“My co-authors and I are currently conducting online lab experiments to explore how humans can improve their prompts and achieve better results with these programs,” Shin says. “Rather than viewing AI as a threat to human intelligence, we should embrace it as a valuable tool that can enhance our capabilities.”