diff --git a/lib/stdlib/src/binary.erl b/lib/stdlib/src/binary.erl
index 9ee1318145af..ef990097bad3 100644
--- a/lib/stdlib/src/binary.erl
+++ b/lib/stdlib/src/binary.erl
@@ -950,8 +950,9 @@ _Example:_
 -doc(#{since => <<"OTP 27.0">>}).
 -spec join([binary()], binary()) -> binary().
 join([], _Separator) -> <<>>;
-join([H], _Separator) when is_binary(H) -> H;
-join([H | T], Separator) ->
+join([H], Separator) when is_binary(H) andalso is_binary(Separator) -> H;
+join([H | T], Separator) when is_binary(Separator) ->
+    %% Starting with an empty binary convinces the compiler to use the new "private append" optimisation
     Acc = <<>>,
     join(T, Separator, <<Acc/binary, H/binary>>);
 join(Arg, Separator) ->
diff --git a/lib/stdlib/test/binary_module_SUITE.erl b/lib/stdlib/test/binary_module_SUITE.erl
index 51e18ae76161..b2d4013f0756 100644
--- a/lib/stdlib/test/binary_module_SUITE.erl
+++ b/lib/stdlib/test/binary_module_SUITE.erl
@@ -261,6 +261,7 @@ badargs(Config) when is_list(Config) ->
     badarg = ?MASK_ERROR(binary:encode_hex(#{})),
     badarg = ?MASK_ERROR(binary:encode_hex(foo)),
 
+    badarg = ?MASK_ERROR(binary:join(<<"">>, ",")),
     badarg = ?MASK_ERROR(binary:join([""], <<",">>)),
     badarg = ?MASK_ERROR(binary:join([123], <<",">>)),
     badarg = ?MASK_ERROR(binary:join(123, <<",">>)),