Great question—this gets to a key assumption behind Minimax and Alpha-Beta pruning, so let’s unpack it clearly.

First, remember that Minimax (and its Alpha-Beta variant) is designed to compute the optimal strategy under the assumption that the opponent will always make the best possible move for themselves (i.e., the worst move for you).

If your opponent is playing random, suboptimal legal moves instead, here’s the deal:

• The strategy your AI computes is still a safe optimal choice. It guarantees that even if the opponent suddenly decides to play optimally at any point, you’ll still get the best possible outcome you can force.
• You won’t get a strategy that actively exploits the opponent’s bad play (like choosing moves that set up traps the opponent is likely to fall for), but your AI’s moves will never be worse than the optimal baseline. In fact, when the opponent makes a bad move, your AI will end up in a better position than it originally anticipated, and its next search will adapt to that new, more favorable state.

Alpha-Beta pruning works by cutting off branches where the current path can’t possibly lead to a better result than what we’ve already found—based on the opponent playing optimally.

If the opponent’s suboptimal move is in a pruned subtree, here’s what happens:

• During your AI’s initial search, that branch was ignored because we assumed the opponent would never choose it (since it’s worse for them). So your AI’s original decision doesn’t account for this move.
• If the opponent actually makes that pruned, suboptimal move, your AI will just run a new Minimax/Alpha-Beta search from the new game state. The prune from the previous search doesn’t affect this new calculation, because we’re now evaluating a fresh position where the opponent has already made their (bad) move.
• There’s no risk of your AI making a bad decision because of the prune, since the prune only discarded paths that the opponent wasn’t supposed to take in the first place. The AI’s core logic still holds when faced with unexpected moves—it just re-calculates from the current state.

内容的提问来源于stack exchange，提问作者J.Doe