Superhuman Artificial Intelligence Can Improve Human Decision Making by Increasing Novelty

Abstract How will superhuman artificial intelligence (AI) affect human decision making? And what will be the mechanisms behind this effect? We address these questions in a domain where AI already exceeds human performance, analyzing more than 5.8 million move decisions made by professional Go players over the past 71 years (1950-2021). To address the first question, we use a superhuman AI program to estimate the quality of human decisions across time, generating 58 billion counterfactual game patterns and comparing the win rates of actual human decisions with those of counterfactual AI decisions. We find that humans began to make significantly better decisions following the advent of superhuman AI. We then examine human players’ strategies across time and find that novel decisions (i.e., previously unobserved moves) occurred more frequently and became associated with higher decision quality after the advent of superhuman AI. Our findings suggest that the development of superhuman AI programs may have prompted human players to break away from traditional strategies and induced them to explore novel moves, which in turn may have improved their decision-making.

Introduction. “It made me question human creativity. When I saw AlphaGo’s moves, I wondered whether the Go moves I ha[d] known were the right ones. Its style was different, and it was such an unusual experience that it took time for me to adjust. AlphaGo made me realize that I must study Go more.” (1) - Sedol Lee, a former world Go champion Recent advances in artificial intelligence (AI) have resulted in automated systems that approach or surpass human performance in fields as diverse as medicine (e.g., diagnosing diseases) (2), transportation (e.g., autonomous driving) (3), language (e.g., ChatGPT based on GPT-3) (4), and natural sciences (e.g., AlphaFold) (5), among others (6). As AI systems outperform humans in these settings, a natural question to ask is how humans will change their own decision-making. Humans will likely adjust to such advancements in AI by delegating to, receiving aid from, or learning from AI systems to improve their own performance. But will human decision-making itself change? And what will be the mechanisms underlying this change?

Answering these questions is challenging because individual human decisions are not necessarily recorded, and AI systems are only gradually being adopted in selective domains. In this article, we overcome these challenges by studying a domain in which detailed records of human decisions are available and where the advent of superhuman AI (7) can be connected to a specific date: March 15, 2016, when AlphaGo—an AI program developed by Google’s DeepMind—shocked the world by defeating a human world champion in Go. We analyze more than 5.8 million decisions made by professional Go players over the past 71 years (1950 to 2021), using a superhuman AI system to evaluate the quality of these decisions. By looking at Accepted Article THE IMPACT OF AI ON DECISION QUALITY AND NOVELTY IN GO 4 The published version of this article is available at PNAS: https://doi.org/10.1073/pnas.2214840120 how human play differed before and after the advent of superhuman AI1; we are able to evaluate its impact.

Our analyses of professional Go players’ decisions reveal that human decision-making significantly improved following the advent of superhuman AI. We also find that this 1 We use the term “advent of superhuman AI” to denote a series of events that occurred between 2016 and 2017, including AlphaGo’s victory over the human world champion in March 2016 and developments of superhuman AI programs between 2016 and 2017. These notable events are listed in SI Appendix, section 1.

Accepted Article THE IMPACT OF AI ON DECISION QUALITY AND NOVELTY IN GO 5 The published version of this article is available at PNAS: https://doi.org/10.1073/pnas.2214840120 improvement may be partly explained by increased novelty in human decisions made after the exogenous shock of this event. These findings illustrate that the development of superhuman AI may result in improvements in human decision-making, with innovations spreading from machines to humans and spurring further novel developments among those humans.

Related work. Questions about the impact of superhuman AI on human behavior are related to the literature on cumulative cultural evolution. This literature shows that there is no guarantee that human decision-making will improve in response to innovations, despite the human ability to accumulate knowledge within and across generations (8, 9). Often, cumulative cultural evolution does occur, as superior forms of decision-making are transferred from one group of individuals to another (10, 11). However, at times, intrinsic biases and frictions in human learning can delay or derail such process (12, 13). When there exist suboptimal but familiar decisions whose efficacy has been demonstrated by others, even experts fail to adopt unfamiliar but objectively better alternatives (14). It is thus not obvious whether human decision-making will improve following advancements in AI.

If innovations produced by superhuman AI do result in changes in human decisionmaking, a second question is what mechanism might underlie this process. Previous research suggests that novelty could be a relevant factor. It has been proposed that AI systems can generate new ideas by combining familiar ideas in novel ways and exploring conceptual spaces that may have been overlooked (15, 16). However, relatively little research has investigated whether human decision-makers will readily adopt ideas generated by an AI system; for an exception, see ref. 17.

Method. We present our main results from three different sets of analyses. First, we estimate the quality of human decisions over time by using KataGo (18), a superhuman AI program. We use it to simulate billions of game patterns (i.e., 10,000 game patterns for each of the 5.8 million decisions) and compare the win rates of actual human decisions with those of counterfactual (optimal) AI decisions to construct a Decision Quality Index for each human decision (hereafter, DQI; Materials and Methods and SI Appendix, section 2.2 for details). Second, we estimate the novelty of human decisions over time by examining move sequences and identifying the first historically novel move of each game. (An additional analysis on novelty estimated from a different measure is presented in SI Appendix, Fig. S1.) Finally, we use our estimates of decision quality and novelty to estimate models testing the hypothesis that AI improved human decision making by encouraging novel decision-making.

Discussion. As AI systems continue to approach or surpass human abilities in various fields, it is essential to comprehend the effects they have on human decision-making (23–26). This research topic, specifically within the context of the game of Go, primarily has centered on assessing the Accepted Article THE IMPACT OF AI ON DECISION QUALITY AND NOVELTY IN GO 17 The published version of this article is available at PNAS: https://doi.org/10.1073/pnas.2214840120 change in human decision quality. The increase in human decision quality following the advent of superhuman AI was initially documented by the first two authors of this article (22) and was subsequently corroborated by other research groups (27). However, to our knowledge, none of this research simultaneously investigated changes in decision novelty and linked them with the increase in decision quality.

In this research, we find that human decision-making significantly improved following the advent of superhuman AI and that this improvement was associated with greater novelty in human decisions. Because AI can identify optimal decisions free of human biases (especially when it is trained via self-play), it can ultimately unearth superior solutions previously neglected by human decision-makers who may be focused on familiar solutions. The discovery of such superior solutions creates opportunities for humans to learn and innovate further.

One important question concerns how much of the observed increase in decision quality and novelty can be attributed to players internalizing the AI systems’ superior decision-making logic as opposed to merely memorizing AI decisions. Our analyses both in the main text and SI Appendix show that memorization of AI decisions cannot be the sole explanation for the increase in decision quality and novelty. For example, when we exclude from our analysis all human decisions that matched the optimal AI decisions and examine the quality of the remaining human decisions, we still find a sharp increase in decision quality after the advent of superhuman AI.

Our findings raise interesting questions for future research, including 1) whether the advent of superhuman AI increased novelty and thereby increased decision quality (i.e., whether each link in the possible causal chain can be established), 2) through which mechanism(s) superhuman AI increased novelty and decision quality (if not through novelty), 3) how the historically novel decisions themselves qualitatively differed before versus after the advent of superhuman AI, and 4) how styles of decision-making changed after the advent of superhuman Accepted Article THE IMPACT OF AI ON DECISION QUALITY AND NOVELTY IN GO 18 The published version of this article is available at PNAS: https://doi.org/10.1073/pnas.2214840120 AI. There are numerous other related questions worth examining, and we hope that our findings can encourage such investigations not only in contexts similar to Go but also in other contexts that allow careful examination of human decision-making. The increasing availability of superhuman AI systems opens exciting new frontiers for studying human cognition (28).