diff --git a/python/katago/game/board.py b/python/katago/game/board.py
index 7c5002986e..09d38a55fd 100644
--- a/python/katago/game/board.py
+++ b/python/katago/game/board.py
@@ -731,7 +731,7 @@ def remove_unsafe(self,group):
 
             #Zero out all the stuff
             self.board[loc] = Board.EMPTY
-            self.zobrist ^= Board.ZOBRIST_STONE[opp][loc]
+            self.zobrist ^= Board.ZOBRIST_STONE[pla][loc]
             self.group_head[loc] = 0
             self.group_next[loc] = 0
             self.group_prev[loc] = 0
diff --git a/python/tests/test_board_zobrist.py b/python/tests/test_board_zobrist.py
new file mode 100644
index 0000000000..282c90744b
--- /dev/null
+++ b/python/tests/test_board_zobrist.py
@@ -0,0 +1,60 @@
+"""
+Unit tests for the incremental zobrist hash in katago.game.board.Board.
+
+Board.zobrist is updated incrementally as stones are placed and captured. These tests check that it always equals
+the hash recomputed from scratch for the stones currently on the board, in particular after captures.
+"""
+
+import random
+
+from katago.game.board import Board
+
+
+def recomputed_zobrist(board):
+    h = 0
+    for loc in range(board.arrsize):
+        if board.board[loc] == Board.BLACK or board.board[loc] == Board.WHITE:
+            h ^= Board.ZOBRIST_STONE[board.board[loc]][loc]
+    return h
+
+
+def test_zobrist_after_single_capture():
+    board = Board(9)
+    board.play(Board.WHITE, board.loc(1, 0))
+    board.play(Board.BLACK, board.loc(0, 0))
+    board.play(Board.BLACK, board.loc(2, 0))
+    board.play(Board.BLACK, board.loc(1, 1))  # captures the white stone at (1,0)
+    assert board.board[board.loc(1, 0)] == Board.EMPTY
+    assert board.zobrist == recomputed_zobrist(board)
+
+
+def test_zobrist_matches_recomputation_in_random_games():
+    rand = random.Random(12345)
+    for _ in range(100):
+        board = Board(5)
+        pla = Board.BLACK
+        for _ in range(80):
+            moves = [loc for loc in range(board.arrsize) if board.board[loc] == Board.EMPTY and board.would_be_legal(pla, loc)]
+            if not moves:
+                break
+            board.play(pla, rand.choice(moves))
+            assert board.zobrist == recomputed_zobrist(board)
+            pla = Board.get_opp(pla)
+
+
+def test_zobrist_restored_by_undo():
+    rand = random.Random(54321)
+    for _ in range(100):
+        board = Board(5)
+        pla = Board.BLACK
+        for _ in range(80):
+            moves = [loc for loc in range(board.arrsize) if board.board[loc] == Board.EMPTY and board.would_be_legal(pla, loc)]
+            if not moves:
+                break
+            zobrist_before = board.zobrist
+            record = board.playRecordedUnsafe(pla, rand.choice(moves))
+            assert board.zobrist == recomputed_zobrist(board)
+            board.undo(record)
+            assert board.zobrist == zobrist_before
+            board.play(pla, rand.choice(moves))
+            pla = Board.get_opp(pla)