AI Clones of Chess Masters Can Predict Their Real Games
Researchers built AI copies of 8 elite players — and the copies played like the humans.
Can you model people separately, then combine those models to reproduce interactions you never showed them? Researchers tested a controlled version of that question through chess: they took 8 elite chess players, sealed their direct pairwise games, learned each player independently using different methods, and then composed the resulting models on the withheld dyads. To evaluate the generated interactions, they used two measurements — opening-family total variation distance and win-draw-loss (WDL) total variation distance. M1 reduced WDL-TV while leaving opening-family TV largely unchanged, whereas M2 substantially reduced opening-family TV while having little effect on WDL-TV. An additional post-hoc method combining components of the other two retained improvements across both measurements. The results suggest that independently learned models can recover aspects of previously unseen interactions, and that recovery across different behavioural aspects need not be mutually exclusive.
- The AI learned each of 8 elite chess players separately, then combined the copies to predict their head-to-head games — which had been hidden from it.
- Two methods worked differently: one nailed the win/loss/draw results, the other matched the players' favourite opening moves; a combined version improved both.
- It's a chess-only experiment, so real-world uses like simulating customers or coworkers are a possibility, not a product.
Why It Matters
Signals a future where AI simulates how people interact — useful for testing decisions before real-world cost.