An amateur Go player beat a top-ranked AI system in 14 of 15 games, using tactics suggested by a computer program that had analyzed the AI system for weaknesses
But a human just trounced one of the world's best Go AIs 14 games to 1: https://arstechnica.com/... … Tweets: @0xabad1dea : the overwhelming moral of the story recently is that AI is terrifyingly dangerous but also hilariously easily tricked https://arstechnica.com/... @defcon : Hooray Team Human, but the article also reminds us that we're the plucky underdog now. At least in ancient board games. https://twitter.com/... https://twitter.com/... @hugospiers : https://arstechnica.com/... I think this news story is more interesting than it might first appear (without knowing details, so grain of salt). It isn't just a gaming curiosity; it points to a fundamental flaw with “deep learning” approaches in (2/3) Artur Ekert / @aekert : “...The tactics used by Pelrine involved slowly stringing together a large “loop” of stones to encircle one of his opponent's own groups, while distracting the AI with moves in other corners of the board...” https://twitter.com/... Matt Willemsen / @mattotcha : Man beats machine at Go in human victory over AI - Amateur exploited weakness in systems that have otherwise dominated grandmasters. | Ars Technica https://arstechnica.com/... @yoavgo : interesting! computer-identified blind-spot in Go engines, that can be exploited by a not-too-strong human player to force wins. but do anyone understand this paragraph? https://twitter.com/... https://twitter.com/... Carl Benedikt Frey / @carlbfrey : Amateur Go player beats cutting-edge AI based on rarely used strategy. Underlines how AI still struggles to learn from small data - a key bottleneck going forward. https://www.ft.com/... Lev Reyzin / @lreyzin : I find this both interesting and unexpected. I did not imagine MCTS would be vulnerable to adversarial attacks. And that a human could implement such an attack during a real game is even more surprising. “Man beats machine at Go in human victory over AI ” https://www.ft.com/... Tae Kim / @firstadopter : A.I. is not bright “A human player has comprehensively defeated a top-ranked AI system at the board game Go, in a surprise reversal of the 2016 computer victory .. systems can “understand” only specific situations they have been exposed to in the past” https://www.ft.com/... Balaji / @balajis : Reminds me of work on adversarial images. Is this an adversarial strategy, where a human/computer combo finds a strategy that beats another computer — but not a human? https://twitter.com/... https://twitter.com/... James Pethokoukis / @jimpethokoukis : This is some real Econ 101 stuff here showing how machines can make humans more productive. 👨 + 🤖 = ⤴️ https://twitter.com/... Dan Elton / @moreisdifferent : 1/ Trying to signal boost w a short 🧵 Amateur Go player Kellin Pelrine can consistently beat “KataGo”, an AI system that was once classified as “strongly superhuman”. Strikingly, the strategies employed to beat the AI do not foil other amateur players. https://www.ft.com/... https://twitter.com/... Amandeep Gill / @gioasempre : Notwithstanding the breathless commentary of the past few years, still time to build a correct understanding of what AI can, and more importantly cannot do. This 👇🏽 https://www.ft.com/... Zachary Kallenborn / @zkallenborn : Important piece, especially for debates on AI in the military. AlphaGo beat the world's best Go players, then an amateur human beat it using simple tricks. Except in war, states will confront an entire intelligence apparatus dedicated to finding the flaws in an autonomous weapon https://twitter.com/... Craig Thorburn / @thorburntoronto : First example I have seen in which a person hacks powerful #AI with a technique that would be obvious to a human. Will not be the last example. https://twitter.com/... Andrew Orlowski / @andreworlowski : This is just kicking the bots when they're down 🫣 https://twitter.com/... John Thornhill / @johnthornhillft : “We've been far too hasty to ascribe superhuman levels of intelligence to machines,” says Prof Stuart Russell https://www.ft.com/...
Context & Ripple Effects
The arc from Deep Blue to AlphaGo had settled into a comfortable narrative: machines win at board games because they simply play better. Kellin Pelrine broke it — an amateur beating top-ranked KataGo 14 games to 1 by playing tactics suggested by a separate program that had probed KataGo and found a blind spot.
This is not a return to human superiority but a new division of labor: the human executes, another algorithm finds the weakness. It echoes what Garry Kasparov argued after his own loss to Deep Blue in his interview on staying ahead of the machines — humans stay competitive when paired with tools, not when playing alone.
First-order effects
- Pelrine's 14-1 result directly punctures KataGo's standing as effectively unbeatable at Go; its published strength no longer guarantees safety against an attacker who knows the exploit.
Second-order effects
- Other game-AI developers now face pressure to run adversarial probing against their own systems before deployment, turning red-teaming from optional research into a quality gate.
Third-order effects
- If the pattern holds beyond Go, the meaningful benchmark for AI shifts from raw performance to robustness under attack — and the winning competitor becomes the human-plus-probing-tool combination rather than either alone.
The trend: Game-playing AI is entering an adversarial phase where a second algorithm that probes for weaknesses lets humans defeat systems once considered insurmountable.