Meridiano

Add Article

No Image Available

Article Details

Title

Article: Figuring out why AIs get flummoxed by some games

Impact Score

5 / 10

AI Summary (Processed Content)

Google DeepMind's Alpha AIs, which master games through self-play, have been found to have exploitable blind spots, as demonstrated by specific Go positions. These failures are not trivial, as analyzing them helps identify critical weaknesses in AI training methods that could impact broader real-world applications.

A new research paper identifies an entire category of "impartial games," like the simple matchstick game Nim, where the self-play training method used for AlphaGo fundamentally fails. In impartial games, players share pieces and rules, unlike in chess.

The research shows that because any impartial game position can be mathematically represented by a Nim configuration, the failure in Nim applies to all games in this category. This reveals a significant limitation in current AI training approaches for these types of strategic problems.

Main Topics: AI game-playing limitations, failure modes in AI training, impartial games (specifically Nim), and the implications of research findings for broader AI reliability.

Original URL

https://arstechnica.com/ai/2026/03/figuring-out-why-ais-get-flummoxed-by-some-games/

Source Feed

Ars Technica

Published Date

2026-03-13 21:47

Fetched Date

2026-03-13 19:30

Processed Date

2026-03-13 19:31

Embedding Status

Present

Cluster ID

Not Clustered

Raw Extracted Content

With its Alpha series of game-playing AIs, Google’s DeepMind group seemed to have found a way for its AIs to tackle any game, mastering games like chess and Go by repeatedly playing itself during training. But then some odd things happened as people started identifying Go positions that would lose against relative newcomers to the game but easily defeat a similar Go-playing AI.
While beating an AI at a board game may seem relatively trivial, it can help us identify failure modes of the AI, or ways in which we can improve their training to avoid having them develop these blind spots in the first place—things that may become critical as people rely on AI input for a growing range of problems.
A recent paper published in Machine Learning describes an entire category of games where the method used to train AlphaGo and AlphaChess fails. The games in question can be remarkably simple, as exemplified by the one the researchers worked with: Nim, which involves two players taking turns removing matchsticks from a pyramid-shaped board until one is left without a legal move.
Impartiality
Nim involves setting up a set of rows of matchsticks, with the top row having a single match, and every row below it having two more than the one above. This creates a pyramid-shaped board. Two players then take turns removing matchsticks from the board, choosing a row and then removing anywhere from one item to the entire contents of the row. The game goes until there are no legal moves left. It’s a simple game that can easily be taught to children.
It also turns out to be a critical example of an entire category of rule sets that define “impartial games.” These differ from something like chess, where each player has their own set of pieces; in impartial games, the two players share the same pieces and are bound by the same set of rules. Nim’s importance stems from a theorem showing that any position in an impartial game can be represented by a configuration of a Nim pyramid. Meaning that if something applies to Nim, it applies to all impartial games.