AlphaGo: The AI That Beat the World Champion at Go

On 9 March 2016, inside a hotel ballroom in Seoul, South Korea, a computer program called AlphaGo sat down — metaphorically — to play the ancient board game of Go against Lee Sedol, the greatest Go player of his generation and arguably the greatest of the modern era. Most experts expected Lee Sedol to win comfortably. Instead, AlphaGo won four of the five games, sending shockwaves through both the AI research community and the wider world. It was, by any measure, one of the most significant milestones in the history of artificial intelligence.

This article explains what Go is, why it was considered uniquely difficult for AI, how DeepMind built AlphaGo, what happened during the match — including the legendary Move 37 — and what the broader legacy of AlphaGo has been for AI research and development.

What Is the Game of Go?

Go is a two-player strategy board game that originated in China more than 2,500 years ago, making it one of the oldest games still played today in its original form. The rules are deceptively simple. Players take turns placing black or white stones on the intersections of a 19×19 grid. The objective is to surround more territory — empty intersections — than your opponent. Stones that are completely surrounded by the opponent’s stones are captured and removed from the board. That is essentially it.

The simplicity of the rules, however, conceals a depth of strategic complexity that no other game can match. Chess has roughly 10^44 possible board positions. Go has approximately 10^170 — a number so astronomically large that it dwarfs the estimated number of atoms in the observable universe (around 10^80). A Go game typically lasts around 150 to 250 moves, and at each move a player may have 200 or more plausible options to choose from.

Go is also unique in that top-level play relies heavily on intuition and pattern recognition developed over decades of study. Professional Go players — who begin training in childhood and study the game with the intensity of Olympic athletes — often cannot fully explain why they make the moves they do. The knowledge is embodied, aesthetic, almost ineffable. This made Go seem like exactly the kind of human activity that AI could not replicate.

You can find a broader overview of AI capabilities in our guide to what artificial intelligence is.

Why Go Was Considered Impossible for AI

The standard technique that had made Deep Blue a world chess champion in 1997 — minimax search with aggressive pruning of the game tree, combined with a hand-crafted evaluation function — simply could not scale to Go. The branching factor in chess is roughly 35; in Go it is roughly 250. Even with the most aggressive pruning, searching to the depths needed to play strong Go would require computational resources far beyond anything available.

Conventional AI researchers had tried applying Monte Carlo Tree Search (MCTS) to Go starting in the mid-2000s, and this produced programs that could play at a reasonable amateur level. But even the best MCTS-based Go programs, such as Zen and Crazy Stone, were nowhere near the level of strong professionals. The fundamental problem was the evaluation function: how do you assess the strength of a given board position in Go? In chess, you can assign approximate values to pieces and count material. In Go, there are no pieces with fixed values, and the significance of a stone depends entirely on the surrounding context in ways that are extraordinarily difficult to formalise.

The conventional wisdom in the AI community, as late as 2014, was that computer Go would not reach the professional level for at least a decade, perhaps much longer. Some researchers doubted it would ever happen with the techniques available at the time. That consensus was about to be spectacularly overturned.

How DeepMind Built AlphaGo

DeepMind, a London-based AI research company founded in 2010 and acquired by Google in 2014, had been working on combining deep learning with reinforcement learning — the branch of AI concerned with training agents to take actions in environments to maximise cumulative reward. The company’s work on learning to play Atari video games directly from pixels had attracted enormous attention in 2013, but Go was a very different and far more challenging problem.

The AlphaGo system, described in a landmark paper published in Nature in January 2016, combined three key innovations. First, a deep convolutional neural network trained on millions of games played by human amateurs and professionals — this ‘policy network’ learned to predict which moves human experts would make in any given position, giving AlphaGo a strong prior over plausible moves. Second, a ‘value network’ — another deep neural network — trained to predict who would win from a given board position, providing the evaluation function that had been so elusive. Third, a Monte Carlo Tree Search algorithm that used both the policy and value networks to guide its search, focusing on the most promising moves and positions rather than brute-forcing through every possibility.

Crucially, DeepMind went beyond simply imitating human play. After training on human games, AlphaGo played millions of games against itself, using reinforcement learning to improve: wins were rewarded, losses were penalised, and the network weights were updated accordingly. This self-play phase allowed AlphaGo to develop strategies that went beyond the human knowledge encoded in the training data.

For more context on the history that led to AlphaGo, see our complete history of artificial intelligence.

The Lee Sedol Match: Game by Game

The five-game match between AlphaGo and Lee Sedol took place in Seoul from 9 to 15 March 2016. It was live-streamed on YouTube and watched by an estimated 200 million people across Asia. The prize fund was one million US dollars, with $900,000 going to Lee Sedol if he won (it was later donated to charity by DeepMind).

Game 1: AlphaGo played with unusual aggression and won decisively. Lee Sedol appeared shaken. He had predicted before the match that he would win 5-0 or at worst 4-1. ‘I was quite shocked,’ he said afterwards. Game 2: AlphaGo continued to play in unexpected ways, making moves that top commentators described as both brilliant and deeply strange. AlphaGo won again. The commentators — professional 9-dan players — were struggling to explain the computer’s strategy. Game 3: Another AlphaGo victory. Lee Sedol was now 0-3 down with the match already lost (the winning threshold was three games). The mood among professional Go players worldwide was sombre.

Game 4, Move 37, and Lee Sedol’s Miracle: Down 0-3, Lee Sedol was playing what many expected to be a consolation game. Then, on Move 78, he played a move — a stone placed in a position that professional commentators described as extraordinary — that AlphaGo, calculating what seemed like the perfect response, failed to handle correctly. AlphaGo had not seen this move coming; its value network had assigned it only a 1-in-10,000 probability. Lee Sedol won Game 4, and the Go world erupted with joy. Even in defeat, the humans had found one move the machine had not learned.

But it is Game 2 that is most remembered by AI researchers, because of Move 37 — a move made by AlphaGo, not Lee Sedol. On the 37th move of Game 2, AlphaGo placed a stone on the fifth row from the edge of the board in a position that top commentators immediately called a mistake. ‘That’s a very strange move,’ said one 9-dan professional watching the game. ‘I thought it was a mistake,’ said another. Lee Sedol left the playing room for fifteen minutes, visibly unsettled. But as the game progressed, it became clear that Move 37 was not a mistake. It was a brilliant positional move that would prove decisive many moves later. AlphaGo had invented a move that human players had not previously considered — a move that expanded the known vocabulary of Go.

AlphaGo Zero and the Leap to Superhuman Intelligence

Following the Lee Sedol match, DeepMind continued to refine and extend the AlphaGo system. AlphaGo Master, an improved version, played 60 online games against top professionals in January 2017 and won every single one. Then, in October 2017, DeepMind published a paper in Nature introducing AlphaGo Zero.

AlphaGo Zero was a fundamental departure from its predecessors. The original AlphaGo had been trained on millions of human games to bootstrap its understanding of the game. AlphaGo Zero started with nothing but the rules of Go and played against itself from scratch, learning entirely through self-play with reinforcement learning. It was given no human knowledge and no human games to learn from.

The results were astonishing. Starting from random play, AlphaGo Zero reached the level of the original AlphaGo — the version that had beaten Lee Sedol — in just 36 hours. After 72 hours it was stronger than AlphaGo Master. Within 40 days it was the strongest Go player in the world, human or machine, by a considerable margin. The system developed entirely novel strategies that differed dramatically from established human Go theory, including rediscovering some classic human joseki (established opening patterns) and discarding others that human players had considered essential.

DeepMind then went further still, publishing AlphaZero in 2018 — a single system trained from scratch using self-play that mastered not just Go but also chess and the Japanese board game shogi, achieving superhuman performance in all three games simultaneously. The approach was confirmed: given clear rules and a reward signal, self-play reinforcement learning combined with deep neural networks could produce superhuman game-playing AI from scratch, without human knowledge.

For more AI concepts explained simply, visit the AI glossary.

Legacy and Impact on AI Research

The impact of AlphaGo on the broader field of AI has been profound and lasting. Before AlphaGo, it was widely assumed that the kind of intuitive, pattern-recognition-based reasoning required for top-level Go — and by implication, for many other complex human activities — was uniquely human and would remain beyond the reach of machines for years or decades. AlphaGo demolished that assumption, not gradually, but all at once.

The techniques developed for AlphaGo — combining deep convolutional networks with reinforcement learning and self-play — have been applied to a wide range of problems outside of games. DeepMind’s AlphaFold, which solved the protein folding problem, drew directly on lessons learned from AlphaGo. OpenAI applied similar self-play reinforcement learning techniques to train AI systems that beat professional human teams at the video game Dota 2. Researchers have applied the approach to robotics, chip design, materials discovery, and more.

AlphaGo also had a significant cultural impact. Professional Go players worldwide went through a period of existential reckoning similar to what chess players experienced after Deep Blue. Some initially retreated from AI, refusing to study its games. Most eventually embraced it as a tool for improving their own play, much as chess players today routinely use engines for analysis. The Go world has been transformed: opening theory has been revolutionised, AI-discovered moves have become mainstream, and the barrier between human and AI play has dissolved.

For the latest generation of AI tools that have followed in AlphaGo’s wake, see our guide to Google Gemini.

Lee Sedol himself retired from professional Go in November 2019, citing AI as one reason for his decision. ‘Even if I become the number one, there is an entity that cannot be defeated,’ he said. He had delivered one of the most remarkable performances in the history of competitive games — winning a single game against a superhuman AI — but he had also understood more clearly than most what the AlphaGo moment really meant.

The Cultural and Philosophical Aftermath

The AlphaGo match did not just shock the professional Go world; it prompted a broader public conversation about what artificial intelligence is, what it can do, and what its rise means for humanity. The five games in Seoul were watched by an estimated 200 million people across Asia — a viewership that dwarfed many major sporting events. Go, already deeply embedded in East Asian culture as a symbol of intellectual achievement and strategic wisdom, had become the stage for a confrontation that felt genuinely historic.

Professional players and commentators initially struggled to process what they had seen. Many had assumed, going into the match, that Lee Sedol — a nine-time world champion known for his creative, aggressive style — would win convincingly. His pre-match prediction of a 5-0 or 4-1 victory was not mere bravado; it reflected the genuine consensus among the top players in the world. When AlphaGo won the first three games without apparent difficulty, there was a period of collective disbelief. Prominent players described it as a ‘tsunami’ that had fundamentally changed their understanding of the game.

The cultural impact was particularly felt in South Korea and China, where Go has historically been seen as a uniquely human domain requiring not just calculation but wisdom, creativity, and understanding of beauty. The question of whether AlphaGo ‘understood’ Go in any meaningful sense — or whether it was merely computing at a superhuman scale — was debated earnestly in newspapers and television programmes. Western philosophy of mind had grappled with similar questions since Alan Turing’s day, but the AlphaGo match made them viscerally immediate and personally relevant for millions of people.

The documentary ‘AlphaGo,’ directed by Greg Kohs and released in 2017, captured the emotional texture of the match and its aftermath with unusual depth. It showed not just the technical achievement but the human stories — Lee Sedol’s composure and ultimate resilience, the DeepMind team’s nervous tension before each game, and the reflections of Korean players and commentators struggling to make sense of what the machines had done. The film won an Emmy Award and has been widely praised as the most compelling account of an AI milestone ever filmed.

AlphaGo’s Influence on Strategy Games and Beyond

The influence of AlphaGo on strategy games has been as profound as its scientific impact. In professional chess, engine analysis has been standard practice since the early 2000s, but the AlphaGo moment prompted Go players to adopt AI analysis tools with an urgency that chess players had not shown since Deep Blue. Programs based on AlphaGo’s techniques — including KataGo, Leela Zero, and ELF OpenGo — became widely available to professional and amateur players alike within a year of the match.

The effect on Go theory has been revolutionary. Centuries of accumulated human wisdom about opening patterns (joseki), mid-game fighting, and endgame technique have been revised, confirmed, refuted, or nuanced by AI analysis. Moves that human players had considered clearly suboptimal for decades have been vindicated; openings that were considered standard have been abandoned; entirely new sequences that no human had previously considered have become part of the professional vocabulary. Top professional players now study AI games as part of their regular training, and tournaments have adopted AI-assisted post-game analysis as standard practice.

The shogi world in Japan had a parallel experience: AI programs using similar techniques reached a level well above the best professional players, and professional shogi has similarly been transformed. In poker, different AI techniques — including counterfactual regret minimisation — produced systems that could beat professional players at heads-up no-limit Texas Hold’em, showing that the combination of deep learning with game-theoretic reasoning could conquer a domain defined by imperfect information and psychological complexity.

More broadly, the AlphaGo achievement accelerated interest in applying deep reinforcement learning to real-world problems. If an AI could develop superhuman intuition in Go purely from self-play, what else might it learn by playing against itself? Robotics researchers have applied similar techniques to training robot arms and locomotion controllers. Chip designers have used reinforcement learning to optimise the placement of components on silicon wafers. Pharmaceutical researchers have applied it to the design of new molecules. The conceptual doors opened by AlphaGo have proven extraordinarily productive.

---

Frequently Asked Questions

---

Related Reading

• The Complete History of Artificial Intelligence
• What Is Artificial Intelligence? A Beginner’s Guide
• The AI Glossary: Every Term Explained Simply
• Google Gemini: The Complete Guide

---

Sources

• Grokipedia
• Stanford HAI AI Index Report

You May Also Like

• What Is Artificial Intelligence
• Best AI Tools for Beginners
• How to Use AI
• AI Tools Directory
• Best Free AI Courses