diff --git a/xla/tsl/framework/type_traits.h b/xla/tsl/framework/type_traits.h
index 31c8fe56a78faa..bb43632b3667f4 100644
--- a/xla/tsl/framework/type_traits.h
+++ b/xla/tsl/framework/type_traits.h
@@ -98,17 +98,22 @@ class numeric_limits<tsl::quint16> : public numeric_limits<tsl::uint16> {};
 template <>
 class numeric_limits<tsl::qint32> : public numeric_limits<tsl::int32> {};
 
-// Specialize is_signed for quantized types.
+// Templates from <type_traits> are not permitted to be specialized by users,
+// and doing so may be ignored or cause an error. As a transitional measure,
+// since tensorflow was specializing is_signed for some time, poison uses by
+// specializing with an empty struct (missing the `value` member).
+//
+// TODO(b/392034954): Delete these invalid specializations.
 template <>
-struct is_signed<tsl::qint8> : public is_signed<tsl::int8> {};
+struct is_signed<tsl::qint8> {};
 template <>
-struct is_signed<tsl::quint8> : public is_signed<tsl::uint8> {};
+struct is_signed<tsl::quint8> {};
 template <>
-struct is_signed<tsl::qint16> : public is_signed<tsl::int16> {};
+struct is_signed<tsl::qint16> {};
 template <>
-struct is_signed<tsl::quint16> : public is_signed<tsl::uint16> {};
+struct is_signed<tsl::quint16> {};
 template <>
-struct is_signed<tsl::qint32> : public is_signed<tsl::int32> {};
+struct is_signed<tsl::qint32> {};
 
 }  // namespace std