Introduction

Chess is sometimes used as an example of one of the first fields to benefit from the existence of artificial intelligence systems with superhuman performance. Both human performance and the enthusiasm surrounding human tournaments increased after the advent of such systems. In my opinion, these points are incorrectly cited as a basis for claims that affect broader fields.

Different products generate interest because of their creators or their content. This distinction is important.

Human chess tournaments are watched far more than machine tournaments, even though machines play better. Are human arts threatened by artificial intelligence? One might think that if the spectator prefers human play to that of machines, then they will prefer human books even if those written by machines are better. But chess and the arts differ in some important aspects. Most spectators do not understand the chess games they watch. The game is the stage on which the story that interests them unfolds: the confrontation between two players. It's the actors in the confrontation who interest them, not the medium.

For a film or a book, it's different: the story is the medium. Readers and moviegoers aren't as interested in the author as they are in the content of the work they consume. It's not for nothing that fictional characters are more famous than their authors, and that actors are often better known than directors.

However, this needs to be qualified: People have a very strong negative bias against artificial authors. But their inability to differentiate them from human authors leaves me skeptical about the future of human art as a viable commercial practice if no action is taken.

Chess A.I. is misaligned

Professionals consult them to prepare their openings before playing in a tournament, and casual players have access to them at the end of each game to review their mistakes and missed moves. AI integration is very advanced in chess. One can imagine that in the future, AI will be just as integrated into the practice of every discipline, and that its practitioners will be able to benefit from the same advantages.

We learn more about the true state of AI in chess when we look at the facts that are missing. (I'm reusing many of the arguments from this post on Go here.) Several new ideas have been discovered by machines, but playing against a computer doesn't allow for lasting improvement. Indeed, the majority of move suggestions that an AI can give a human are not usable in practice. Thus, despite the abundant access to these systems, there is no record of a player reaching a competitive level without practice in a club or with a coach. In chess, it is possible to lose the game for reasons that require very little depth to find. For example, moving your queen to a position where it can be captured by a pawn on the next move by your opponent. Machines can correctly identify and inform beginners of these errors. Beyond a certain level, humans no longer make these mistakes, and the game becomes a balance between tactics ^[1] and strategy^[2].

Chess machines are not aligned: they find the moves that allow them to win most easily, not the move that allows a human to win most easily. Their tactics rely on overly long sequences, and their strategies on overly complex concepts.

The most glaring difference is in computation: a move is considered bad by the machine if it has a refutation at depth 48, even though a human could never see it. But this problem affects all aspects of the game in which humans and computers differ. In chess, the hand that moves the pieces must be the same as the one that presses the clock. Consequently, players with limited time move pieces closer to the clock. This situation is impossible to explain to current chess AIs. Their inability to simulate a player other than themselves makes alignment impossible. Work to mitigate this problem is just beginning with Maiachess, a neural network trained to mimic the playing style of a player at a given level. If the future of humanity were at stake, this would not be a sufficient solution.

When resources are limited, the best performance in a field is not a symbiosis between human and machine.

It is sometimes said that the human-machine combination is superior to the performance of a machine alone. This is called centaur chess. This claim is difficult to refute due to a lack of data. However, several factors contradict this. If we compare the graph of machine chess performance over time with the graph of human performance over time, we realize that machines progress faster than humans. As with LLMs^[3], many people don't realize the progress of machines in a field until they surpass them. Furthermore, the fact that assisted amateurs perform better than assisted professional players suggests that machines do not allow for a multiplication of one's chess skills: The players who benefit most from the help of artificial intelligence are those who do not try to understand its moves. Watching computer games analyzed by humans, I get the impression that humans can understand each move played, but that it requires so much time and energy that this style of play wouldn't be viable in real match conditions. Again, knowing how to manage these resources is one of the necessary skills to consider when choosing a move, and machines can't model it.

Conclusion

Chess performance is measurable, and machines have thus achieved a very high level of understanding of the domain. However, these highly advanced machines are not the most suitable for helping humans integrate their knowledge into their practice. Indeed, their inability to model human behavior prevents them from providing results appropriate to the situations a player will encounter. Thus, they seem to be at odds with LLMs: the evaluations that LLMs achieve are constantly challenged as not measuring intelligence, even though they are sometimes able to explain their reasoning when solving a problem.

I believe, however, that chess, due to the maturity of its integration with machines, remains a useful domain for predicting the effects of superhuman systems on certain fields. Programming is the area where these systems are becoming more widespread and developing most rapidly right now, thanks to LLMs. Even if LLMs can explain their reasoning, understanding a complex idea takes time for a human, even if it is very well explained. In competitive production environments, this time spent understanding a system is unacceptable if it can be eliminated. Centaur chess tournaments, with their limited time to play a move, illustrate these constraints and suggest that the most successful programmers will not necessarily be those who can program best without the help of an LLM. If these programmers reach a superhuman level of programming skill, the advantage of being able to efficiently sift through large amounts of generated code using techniques akin to statistical analysis outweighs a deep understanding of the system.

A second aspect of the impact of machines on chess is social. People still prefer to watch humans play despite their suboptimal performance. The stars endure. It's worth noting that there are no "star" programmers in the eyes of the general public, unlike in the arts, culture, and entertainment industries. Even when reducing our model to these categories, I maintain that we need to define whether viewers are primarily interested in the subject matter itself or in its creators to determine if it is "protected" from automation. A Marvel movie is an example of a "product-oriented" domain (who knows the name of the director of their favorite Marvel film?), whereas a soccer match is "author-oriented" (the proportion of people who know the name of a soccer player on their favorite team is comparatively much higher). This is the most useful model I derive from this reflection.

This test of "do people tend to know the names of top performers?" helps explain certain phenomena I didn't understand before. For example, I was surprised by the lack of enthusiasm surrounding artificial pornography. I would have predicted that it would be much more prevalent today, given the image and video generation capabilities of the models. It would seem we have a strong preference for masturbation on human beings, and that regulations implemented by companies with the best models are working. A quick look at r34 for scientific purposes confirms that those who don't mind masturbating to real people or images (which are less regulated) rather than videos have fully embraced AI.

Finally, the role of regulation is paramount. It seems important to me to find a way to allow consumers to know with certainty that they are consuming "organic" content. In artistic circles, some believe they will be saved by a miracle lawsuit that will ban all artificial intelligences trained on copyrighted material. Take the example of music labels, which manage their artists' rights and have real influence on legislation. It's worth remembering that they do this for profit. Since a human costs more than a robot, labels have no interest in protecting their artists if they can partner with companies that artificially generate music, even though consumers would prefer to listen to humans.

1. ^{^}

   Tactics: an advantage determined by calculation. The challenge is to think more deeply than your opponent without making a mistake. For example, sacrificing your queen to expose the opponent's king to a forced sequence of moves that leads to checkmate.

2. ^{^}

   Strategy: an advantage determined by the position of the pieces relative to each other. For example, a pawn will be more valuable if it is near the end of the board because it will become a queen if it can be moved forward.

3. ^{^}

   Large Language Models (LLMs) are the technology behind ChatGPT. They are a type of artificial neural network. The best chess AIs today are all artificial neural networks, but not LLMs, although they can also play the game under certain mysterious conditions.