In 1997, world chess champion Garry Kasparov lost for the first time in history to a computer, Deep Blue. Twenty-seven years later, what has the human defeat against the machine taught us, and can these lessons shed light on the massive arrival of AI in our lives?
Recent advancements in
artificial intelligence (AI), such as the development of generative AIs with the emergence of ChatGPT in November 2022, have raised many questions, hopes, and concerns.
In the spring of 2023, the
US Congress held a hearing with OpenAI, the company that developed ChatGPT, and the European Union has just adopted
its first legislative text regarding AI.
In parliaments and on social media, the rapid advances in AI are fueling discussions. In the future, what impact should we expect on our society? To try to answer this question dispassionately, we suggest looking at what has happened in a field that has already experienced the arrival and victory of AI over human capabilities:
chess. Machines have indeed reached a higher level than humans in this domain for over a quarter of a century now.
Why use chess as an indicator?
Since the early days of
computing, chess has been used as an indicator of software and hardware progress. It's an interesting game to study AI's impacts on society for several reasons:
- It's an intellectual activity that requires various skills: spatial visualization, memory, mental calculation, creativity, adaptability, and so on, skills in which AI competes with human intelligence.
- The game hasn't changed for centuries. The rules are well-established, providing a stable base to study the evolution of players.
- It's possible to objectively measure machine strength and compare this level to that of humans using the
Elo rating system.
- The field of study is limited: it's clear that chess is just a small aspect of life, but that's exactly the point. The narrowness of the topic allows a better focus on the impacts of AI on everyday life.
- AIs have surpassed the level of the best human players for over 20 years. It's therefore possible to examine the concrete impacts on the game of chess and the life of its community, which can be seen as a microcosm of society. We can also study these impacts in light of AI's progression over time.
Let's explore the developments in the chess world since Garry Kasparov, then the reigning world champion, lost a
game against Deep Blue in 1996, followed by the rematch in 1997. We will review several themes that come up in discussions about the risks associated with AI and see how these speculations have played out in the specific field of chess.
Will AI's performances continue to improve ever faster?
There are two major approaches to programming chess software: for a long time, brute force was the only effective method. This involved essentially calculating as quickly as possible to have a deeper move tree, that is, one capable of forecasting the game further into the future.
From an initial position, the computer calculates a series of possibilities to a certain depth, meaning a number of future moves in the game.
Chris Butner, CC BY-SA Today, brute force is competing with AI techniques based on neural networks. In 2018, Google's subsidiary DeepMind produced AlphaZero, an artificial neural network-based deep learning AI that learned chess by playing against itself.
Among the most powerful chess engines today, it's remarkable that both LC0, which is a neural network AI, and Stockfish, which is essentially a brute force calculation engine, achieve comparable results. In the latest ranking of the
Swedish Chess Computer Association (SSDF), they are separated by only 4 Elo points: 3,582 for LC0 versus 3,586 for Stockfish. These two completely different ways of implementing a chess engine are virtually indistinguishable in terms of strength.
In terms of Elo rating, machine progression has been linear. The following chart shows the rating of the strongest software each year according to the SSDF, which began in the mid-1980s. The current leading software, LC0, stands at 3,586, extending the trend as expected.
This linear progression actually reflects a relatively slow advancement in the software itself. In contrast, the increase in computing power has been exponential. This is the famous
Moore's law that states the computing power of computers doubles every eighteen months.
However, Ken Thompson, an American computer scientist who worked in the 1980s on Belle, which was the best chess program at the time, experimentally observed that an exponential increase in computing power leads to a linear increase in software strength, as reflected over the past few decades.
Indeed, adding just one move to the depth of analysis requires calculating many new positions. This reveals that the tree of possible moves becomes increasingly wider with every additional move.
Thus, AI's internal progress seems modest: even without any inherent AI improvements, software strength would still appear to increase solely due to the improvement in machines' computing power. Therefore, we can't attribute all the credit for the constant improvement of chess engines solely to AI's progress.
The reception by the chess community
With the advent of powerful machines in the chess world, the community has inevitably evolved. This aspect is less scientific but perhaps the most important. Let's observe these evolutions.
"Why would people continue playing chess?" This question was genuinely raised right after Kasparov's defeat, as the future of both amateur and professional chess seemed bleak. It turns out that humans prefer playing against other humans and are still interested in the spectacle of strong grandmasters competing with one another, even though machines can identify their mistakes in real time. The prestige of top-level chess players hasn't diminished despite machines being able to beat them.
The style of play has been impacted at many levels. Essentially, players realized there were far more possible approaches to the game than previously thought. Rigid, orthodox rules were dealt a blow. However, it's still essential to analyze the moves made by the machines. AI is also excellent at detecting tactical mistakes, meaning calculation errors over short sequences.
Online, it's possible to analyze games almost instantly. It's somewhat like having a personal tutor at arm's reach. This has surely contributed to a general increase in the skill level of human players and the democratization of the game in recent years.
For now, AIs aren't able to offer good advice on strategy, which involves longer-term considerations within the game. That might change with language models like ChatGPT.
AI has also introduced the possibility of cheating. There have been numerous scandals regarding this, and it must be acknowledged that, to date, there is no "perfect solution" to handle this issue, which mirrors the challenges faced by teachers unsure of whether assignments submitted by students were completed by the students themselves or by ChatGPT.
Preliminary conclusions
This quick overview suggests that, for now, most fears expressed about AI aren't substantiated by experimental evidence. Chess provides an interesting historical precedent to study the impacts of these new technologies when their abilities begin to surpass human capabilities.
Of course, this example is very limited, and it can't be generalized to society as a whole without caution. In particular, chess-playing AIs are not generative AIs like ChatGPT, which have garnered the most attention recently.
Nonetheless, chess is a concrete example that can be useful in putting into perspective the risks associated with AI and the significant influence these systems promise to have on society.