The Suboptimal Move

Bloomberg reports that chess grandmasters are finding new ways to win by playing less optimal moves — deliberately choosing lines that AI engines don't prioritize, because their AI-assisted opponents won't have prepared for them. The strategy sounds paradoxical. It's actually the culmination of a 29-year arc in which humans and AI transformed each other's play — and the latest phase reverses everything that came before it.

The Five Phases

The history of AI in games follows a clean arc. Each phase seemed like an ending. Each turned out to be a transition.

• 1997
  IBM's Deep Blue defeats Garry Kasparov. The first time a machine beats a reigning world chess champion. Kasparov later called it the beginning of a new relationship between humans and machines.
• 2005-2015 The "centaur" era. Human-AI teams beat both pure humans and pure AI. In freestyle tournaments, amateurs with laptops beat grandmasters and supercomputers. The best player isn't the strongest human or the strongest computer — it's a mediocre human with a good process for using a mediocre computer. For a decade, this looks like the permanent equilibrium.
• DEC 2018
  DeepMind's AlphaZero teaches itself to play chess from scratch — no human games, no opening books — and crushes Stockfish, the world's best traditional engine. The centaur era collapses. AI alone is now better than human+AI. The human contribution, which for a decade was the secret ingredient, becomes a liability.
• 2019-2020 Researchers build AI that learns from analyzing human games rather than self-play. DeepMind collaborates with former world champion Vladimir Kramnik to explore new chess variants. AI and humans become co-investigators.
• 2022
  Neural network engines redefine chess creativity. Moves once dismissed as bad are shown to be strong. The game enters a renaissance — more creative, more unconventional, richer than the pre-AI era.

By 2022, the conventional wisdom was settled: AI had solved chess at the highest levels, and humans could only benefit by studying AI's insights. The frontier of the game was wherever AlphaZero and its successors pointed. Every serious player prepared with engines. Every opening was mapped. Every endgame was tabulated.

Then the arc took an unexpected turn.

The Exploit

In February 2023, a story emerged from Go — not chess — that changed the framework. An amateur player beat a top-ranked AI system in 14 out of 15 games. Not by being better at Go. By exploiting a structural blind spot in the AI's evaluation function — feeding it positions where its confidence was high but its assessment was wrong.

The result was treated as a curiosity — a clever hack against a specific Go engine. It was actually a proof of concept. AI systems that evaluate positions by pattern recognition have systematic weaknesses: positions that fall outside their training distribution, sequences that exploit their confidence calibration, lines that look suboptimal by standard metrics but create downstream complications the engine underweights.

An amateur had discovered something profound: AI's strength — exhaustive pattern evaluation — was also its vulnerability. It was only a matter of time before the world's best human players applied the same insight at the highest level.

The Anti-Computer

Three years later, Bloomberg reports that chess grandmasters are doing exactly this. The strategy is called "anti-computer" chess, and it works because of a structural paradox: when every player prepares with the same engine, the engine's recommendations become common knowledge. If you play the engine's top move, your opponent has already studied the engine's top response. The game is pre-analyzed before it starts.

The solution is to step off the analyzed path. Play the engine's third-best move. Or fifth-best. Choose a line that's evaluationally inferior — maybe -0.3 by Stockfish — but practically superior because your opponent hasn't prepared for it. You sacrifice a fraction of theoretical advantage in exchange for an enormous practical advantage: your opponent is now in unfamiliar territory, while you've studied this exact position.

When everyone has access to the same AI, the competitive advantage shifts from who can be most optimal to who can be most unexpectedly suboptimal.

This isn't a bug in chess or a trick that only works once. It's a structural consequence of universal AI access. The moment every player has the same tool, the tool's recommendations become the baseline, not the edge. The edge migrates to wherever the tool doesn't look.

The Pattern

The 29-year arc tells a story in five acts:

Each phase seemed final. Deep Blue's victory was supposed to end competitive chess. The centaur era was supposed to be the permanent equilibrium. AlphaZero was supposed to make human insight irrelevant. The creativity renaissance was supposed to be the mature state.

Instead, each phase created the conditions for the next. AI mastery created universal AI access. Universal access created preparation convergence. Preparation convergence created a premium on unpredictability. And unpredictability — the most human quality — is what the grandmasters are now weaponizing.

Beyond the Board

Chess grandmasters may be the first professionals to complete this cycle, but the structure applies anywhere AI tools become universal. Poker players face a version of this problem — AI-generated optimal strategies are available to everyone, so the edge shifts to reads and timing that exploit opponents' AI-informed patterns. Cybersecurity researchers at DEF CON exploit AI systems by feeding them inputs outside their training distribution — the same structural vulnerability the Go amateur discovered. Employers are returning to in-person interviews because when both sides use AI to optimize resumes and screening, physical presence becomes the only unfakeable signal.

The grandmasters are not Luddites. They use the same engines as everyone else to prepare. They're not rejecting AI — they're rejecting the assumption that the best human strategy is to play whatever the AI recommends. The best human strategy is to play what the AI doesn't recommend — what it rates as slightly worse — because that's where the prepared opponent is weakest.

In 1997, the question was whether AI could beat humans. In 2026, the question has inverted: not whether humans can beat AI, but whether humans can use AI's own certainty against it. The grandmasters say yes. The cost of winning is accepting that the best move isn't always the best move.