Go Players Disempower Themselves to AI: How Cheating Became Undetectable

The LessWrong article "How Go Players Disempower Themselves to AI" by Ashe Vazquez Nuñez examines the sociology of AI cheating in Go tournaments, focusing on a specific case from European team championships. Key points extracted:

Background: AlphaGo vs Lee Sedol

In March 2016, Google DeepMind's AlphaGo defeated world champion Lee Sedol 4-1. Initially, Go culture appeared to adapt similarly to Chess, where AI is used for commentary and teaching without undermining human competition. However, cheating quietly emerged.

The Carlo Metta Case

• Timeline: In early 2018, the European Team Championship's referee accused Italian player Carlo Metta of using AI during an online match.
• AI used: Leela 0.11 (available before AlphaGo paper engines) and later Leela Zero (superhuman open-source engine).
• Evidence: Accusers claimed his move choices matched Leela 0.11 too closely, with a significant disparity between his online play (AI-like) and over-the-board (OTB) play (human-level). However, evidence was presented poorly in a Facebook thread.
• Outcome: Metta was initially banned, but after an appeal by the Italian team, he was exonerated due to circumstantial evidence and community backlash against the stigma of AI cheating accusations.
• Aftermath: Metta's OTB level stagnated, but his online performance skyrocketed: a 9-0 streak in 2018/2019, 9-1 in 2019/2020, and 25 of 26 games won in subsequent years. His only loss was under camera surveillance. The article notes it's now barely disputable among non-Italian European Go players that Metta used AI regularly since 2018.

How Cheating Became Trivial

• The public vilification of AI cheaters backfired, making accusations socially costly.
• The Metta case set a precedent that even obvious cheating could go unpunished through political pressure.
• As a result, online European events saw near-complete lack of enforcement, making cheating trivially easy.

Developers building AI integrity tools for competitive games should note how adversarial dynamics can make detection and punishment ineffective. The sociological factors — stigma, precedent, and organizational inertia — are as important as technical detection methods.