Control territory and capture stones
Go is played by placing stones on the intersections of a grid — 9x9, 13x13 or 19x19 here, chosen with the size buttons below the board, separately from the four AI levels. You place the light stones and move first; the AI plays dark. A stone or connected group whose every neighbouring point is filled by the enemy has no liberties left and is captured. You cannot place a stone where it would have no liberties unless that move captures, and under the ko rule you cannot immediately retake a single stone that has just taken one of yours.
When both players pass in succession the game is scored by area: every stone you have on the board plus every empty point in regions touched only by your stones. The AI receives 7.5 points of komi for moving second, so there are no ties. Dead stones are removed automatically before counting: the game plays out a few hundred quick random continuations, and groups that are captured in most of them and have no two clear eyes are crossed out. It is a heuristic, so capture any stone you are unsure about before passing. The AI passes once nothing useful is left, or when you pass while it is already ahead.
Go on 9x9, 13x13 or 19x19, with the board size chosen separately from the four AI levels. You play the light stones and move first, and the AI receives 7.5 points of komi in return. The core rules are all here: capture by removing a group's last liberty, no suicide unless the move captures, and simple ko, so a lone stone that has just captured a single stone cannot be retaken until a move or pass has been played elsewhere. Scoring is by area, counting your stones on the board plus empty points surrounded only by your stones. Dead stones are removed automatically at the end by a playout-based estimate, and the board shows which stones were removed and which points counted.
The AI first builds a list of sensible moves: it never fills its own eyes, avoids putting its own groups of two or more stones into atari, and ignores small sealed areas unless a capture is at stake. Easy and Medium score each candidate once, rewarding captures, rescues of groups in atari, ataris that actually work, liberties and good opening lines. Hard and AI Training run hundreds or thousands of quick random games from each of their best candidates and play the move that wins most often. The AI passes when nothing useful is left, or when you have passed and its own count says it is already ahead, so games end cleanly after two passes in a row.
Easy uses the one-move scoring with plenty of randomness and no ladder reading, and often settles for any of its top few moves. Medium uses the same scoring without the noise and reads ladders and simple chases before it runs with a stone or attacks one. Hard adds Monte Carlo playouts for about a second per move and is clearly stronger than Medium on 9x9 and 13x13. On 19x19 a second of playouts is spread too thin to add much, and it plays at roughly Medium strength there. AI Training gives the playouts about 1.35 seconds per move on every board, considers more candidate moves, and also learns from every move each playout makes (all-moves-as-first statistics), which sharpens its estimates when playouts are scarce. In testing it beat Hard 32-16 on 9x9 and 10-2 on 13x13; on 19x19 it is still limited by how thinly the playouts are spread.
For the complete rules, the deeper strategy and the history of the game, read the full Go guide, or browse guides for all 33 games.