AI puzzles | The Spectator Australia
Generative AI can paint a thousand dreams with a prompt. Ask it for a cephalopod twirling crockery in a sandstorm and it will oblige without blinking. Yet when you look past the novelty, the seams often show. Nowhere is that clearer than in chess: renderings of people at the board frequently feature impossible positions, miscolored squares, and layouts that seem borrowed from an Escher sketch. Making something that’s not only coherent but genuinely artistic? That’s the real trick.
From surreal images to precise problems
That’s why an emerging strand of research is so intriguing: AI that composes chess puzzles designed to be beautiful, surprising, and sound. A cross-disciplinary team trained a model on a large public trove of tactical positions and then taught it to value the intangibles that human composers prize — clarity, thematic unity, paradox, and the thrill of a bold sacrifice. By scoring candidate puzzles with an “aesthetic sense” and refining via reinforcement learning, the system began to produce new problems with recognizable style. They’re not ready to dethrone the great human composers, but the best results show ideas that feel fresh rather than merely functional.
Puzzle 1: burning the furniture to light the way
Consider one standout example. At a glance, White looks spoiled for choice; rooks loom, the queen eyes entry squares, and Black’s king appears boxed in. Yet direct tries fizzle. For instance, 1 Qa2 runs into 1…Qd4+ with a flurry of checks, and 1 Qa1 (aiming for Rf6–Rg6+) gets snuffed out by 1…Qxg5. The breakthrough requires audacity: throw both rooks into the fire to clear a path for the pieces that matter.
White plays 1 Rg6+!! and if 1…hxg6, then 2 Qa1! Kxf7 3 Qf6+ Kg8 4 Bh6, when the only desperate resource is 4…Qh4+ to postpone 5 Qg7 mate, but the position is strategically busted and the finish is routine. On 1…Kxf7 instead, 2 Qa1 reaches the same winning net. It’s the kind of line that delights judges: violence at the start, restraint in the middle, and a quiet move that ties the bow.
Puzzle 2: the classic smother, with a hook
Another generated composition riffs on the evergreen queen-sacrifice smothered mate. Strip away a pair of rooks from the setup and you’d expect the textbook pattern: 1 Nf7+ Kg8 2 Nh6+ Kh8 3 Qg8+! Rxg8 4 Nf7#, a mating picture many tactics lovers know by heart. But in the actual position those extra rooks change everything. After 1 Nf7+ Black suddenly has 1…Rxf7, and if White tries 2 Rc8+ then 2…Rf8 erects a brick wall. The familiar path is shut.
What makes the construction clever is how it nudges you toward the reflex — the knight checks — and then forces a rethink. You need a preparatory idea that either deflects the defender or alters the move order so the smothered pattern reappears on your terms. It’s not about brute force; it’s about choreography, a tiny twist that transforms a well-worn theme into something that feels newly minted.
Why this matters for chess — and AI
Composing chess puzzles is closer to writing a short poem than solving a math problem. The solution must be unique and correct, but elegance is the point: a sacrificial spark, a tidy geometry of moves, a surprise that makes the final checkmate feel inevitable in retrospect. That AI can learn to chase those qualities — not just legality and tactics, but style — is a meaningful step. The model’s output is already coherent and sometimes charming, especially when it stages quiet intermezzos or uses clearance to set a trap two moves later.
Still, there’s daylight between “interesting” and “immortal.” Human composers draw on an intuition for narrative and novelty, consciously avoiding clichés and remixing known ideas into cleaner, leaner forms. AI, trained on what exists, has a gravitational pull toward the familiar. That’s not a flaw so much as a starting line. Give it better taste metrics, more explicit rules of composition theory, and tighter reinforcement signals, and it could become a powerful co-author — a generator of promising seeds that a human polishes into gems.
For players, the benefits are immediate. Fresh tactical exercises sharpen calculation and pattern recognition; curated aesthetic puzzles rekindle the joy that first hooked many of us on the game. For AI researchers, these micro-worlds are testbeds for creativity under constraints: produce something correct, surprising, and beautiful, or try again.
We’re a long way from AI writing the next great chess anthology on its own. But when a model is brave enough to toss both rooks for a quiet queen slide, or to bend a classic smothered mate into a new silhouette, you can feel the gap narrowing. The magic isn’t that machines can make puzzles. It’s that they’re starting to make ones we actually want to solve.
