diff --git a/numpy/_core/src/multiarray/array_coercion.c b/numpy/_core/src/multiarray/array_coercion.c
index 1b36deca95c2..51aa874bf934 100644
--- a/numpy/_core/src/multiarray/array_coercion.c
+++ b/numpy/_core/src/multiarray/array_coercion.c
@@ -615,10 +615,13 @@ update_shape(int curr_ndim, int *max_ndim,
     return success;
 }
 
-
+#ifndef Py_GIL_DISABLED
 #define COERCION_CACHE_CACHE_SIZE 5
 static int _coercion_cache_num = 0;
 static coercion_cache_obj *_coercion_cache_cache[COERCION_CACHE_CACHE_SIZE];
+#else
+#define COERCION_CACHE_CACHE_SIZE 0
+#endif
 
 /*
  * Steals a reference to the object.
@@ -629,11 +632,14 @@ npy_new_coercion_cache(
         coercion_cache_obj ***next_ptr, int ndim)
 {
     coercion_cache_obj *cache;
+#if COERCION_CACHE_CACHE_SIZE > 0
     if (_coercion_cache_num > 0) {
         _coercion_cache_num--;
         cache = _coercion_cache_cache[_coercion_cache_num];
     }
-    else {
+    else
+#endif
+    {
         cache = PyMem_Malloc(sizeof(coercion_cache_obj));
     }
     if (cache == NULL) {
@@ -662,11 +668,14 @@ npy_unlink_coercion_cache(coercion_cache_obj *current)
 {
     coercion_cache_obj *next = current->next;
     Py_DECREF(current->arr_or_sequence);
+#if COERCION_CACHE_CACHE_SIZE > 0
     if (_coercion_cache_num < COERCION_CACHE_CACHE_SIZE) {
         _coercion_cache_cache[_coercion_cache_num] = current;
         _coercion_cache_num++;
     }
-    else {
+    else
+#endif
+    {
         PyMem_Free(current);
     }
     return next;
diff --git a/numpy/_core/tests/test_multithreading.py b/numpy/_core/tests/test_multithreading.py
new file mode 100644
index 000000000000..8999a18a39ff
--- /dev/null
+++ b/numpy/_core/tests/test_multithreading.py
@@ -0,0 +1,21 @@
+import concurrent.futures
+
+import numpy as np
+import pytest
+
+from numpy.testing import IS_WASM
+
+if IS_WASM:
+    pytest.skip(allow_module_level=True, reason="no threading support in wasm")
+
+
+def test_parallel_errstate_creation():
+    # if the coercion cache is enabled and not thread-safe, creating
+    # RandomState instances simultaneously leads to a data race
+    def func(seed):
+        np.random.RandomState(seed)
+
+    with concurrent.futures.ThreadPoolExecutor(max_workers=8) as tpe:
+        futures = [tpe.submit(func, i) for i in range(500)]
+        for f in futures:
+            f.result()