Researchers developed AI that learned chess not by playing but analyzing reactions of expert commentators in text form to evaluate the quality of the moves
Machines that appreciate “brilliant” and “dumb” chess moves could learn to play the game—and do other things—more efficiently. Tweets: @techreview , @bobehayes , and @kasparov63 Tweets: @techreview : It turns out “brilliant” and “dumb” chess moves could help artificial intelligence learn to play the game—and do other things—more efficiently. https://www.technologyreview.com/ ... Bob E. Hayes / @bobehayes : Instead of practicing, this #AI mastered chess by reading about it https://www.technologyreview.com/ ... “It evaluates the quality of chess moves by analyzing the reaction of expert commentators.” #artificialintelligence #MachineLearning Garry Kasparov / @kasparov63 : Interesting, combining chess & language AI analysis. Always nice to see chess continue its historical role as the drosophila of reasoning, both human & machine! My own article on the theme: https://science.sciencemag.org/ ... https://twitter.com/...
Context & Ripple Effects
This 2019 result inverted the standard recipe: instead of self-play at scale, the system learned move quality from the words experts used about moves — 'brilliant' versus 'dumb' — making human commentary itself the training signal. It landed between two eras of chess AI covered in the related reporting: DeepMind's work with Vladimir Kramnik to have AlphaZero explore new variants, and the moment when [[a:982947|neural-network engines had so redefined the game's creativity that top players turned to deception and misdirection]].
The throughline matters because every later development in the corpus pushes the same direction: away from pure machine optimality and toward hybrid human-machine play, culminating in grandmasters deliberately choosing suboptimal moves to stay competitive.
First-order effects
- Researchers gain a far cheaper path to strong evaluation: mining existing expert text instead of running exhaustive game simulation, which lowers the resource bar for domains where human judgment is already written down.
- Chess commentary acquires new value as structured data — the evaluative language professionals produce becomes a direct input to machine learning rather than mere spectator color.
Second-order effects
- Once human judgment is a usable signal, tools that probe machine weaknesses follow — the same logic behind the [[a:836329|amateur Go player who beat a top-ranked AI in 14 of 15 games using tactics surfaced by a program that had analyzed the system]].
- Engines trained toward perfect play force human competitors to adapt strategically, which is exactly the shift captured by [[a:1166094|grandmasters finding wins in less-than-optimal moves after AI pushed classical chess toward perfection]].
Third-order effects
- If commentary-derived evaluation scales beyond chess, expert-annotated fields — sports, medicine, markets — become training corpora, and the scarcity economics of labeled data shifts toward whoever owns the commentary layer.
- The longer pattern across this coverage is co-evolution rather than replacement: machines push play toward optimality, humans mine the gap between optimal and exploitable, and each side's improvement loop depends on the other.
The trend: Chess AI is moving from computing perfection against humans to ingesting human judgment, making expert players and their language part of the training pipeline rather than its benchmark.