Spending 10 months building the most human-like chess bot

A developer spent ten months building Chessbook-1, a chess bot engineered to behave like a human opponent at any Lichess rating and time control. It was trained on games through July 2026 and evaluated on two September 2026 Lichess datasets: 30,000 random rated positions (all time controls) and 28,000 rated blitz positions (4,000 per rating band). Each model received the position, player ratings and any supported clock/history features, then produced a probability distribution over legal moves; performance was measured by top-1 accuracy and the probability placed on the actual human move. Chessbook-1 reads clocks and move history, Maia-3 reads history but not clocks, and Allie 2.0 reads both but is much larger.

Chessbook-1 predicts human moves strongly, matches Maia-3 overall and beats it where Maia’s training is out-of-distribution (bullet and time scrambles), and effectively ties Allie 2.0 on blitz (pooled top-1: Allie 57.5%, Maia 57.3%, Chessbook-1 57.3%; probability on human ~27-28%). Crucially, Chessbook-1 is the only model calibrated to play at the intended Elo across bands: its average centipawn loss and blunder-rate differences versus humans stay near zero (within about ±1 cp and ±0.6 percentage points), while Maia and Allie over-perform at low Elo and underperform at high Elo. Chessbook-1 also predicts human think time and deliberately errs in time scrambles like humans.