diff --git a/unix/uxmisc.c b/unix/uxmisc.c
index 50d385b6..04af45d0 100644
--- a/unix/uxmisc.c
+++ b/unix/uxmisc.c
@@ -352,11 +352,36 @@ char *make_dir_path(const char *path, mode_t mode)
 
 int open_for_write_would_lose_data(const Filename *fn)
 {
-    FILE *fp;
-    if ((fp = f_open(fn, "r", FALSE)) != NULL) {
-        fclose(fp);
-        return TRUE;
-    } else {
+    struct stat st;
+
+    if (stat(fn->path, &st) < 0) {
+        /*
+         * If the file doesn't even exist, we obviously want to return
+         * false. If we failed to stat it for any other reason,
+         * ignoring the precise error code and returning false still
+         * doesn't seem too unreasonable, because then we'll try to
+         * open the file for writing and report _that_ error, which is
+         * likely to be more to the point.
+         */
         return FALSE;
     }
+
+    /*
+     * OK, something exists at this pathname and we've found out
+     * something about it. But an open-for-write will only
+     * destructively truncate it if it's a regular file with nonzero
+     * size. If it's empty, or some other kind of special thing like a
+     * character device (e.g. /dev/tty) or a named pipe, then opening
+     * it for write is already non-destructive and it's pointless and
+     * annoying to warn about it just because the same file can be
+     * opened for reading. (Indeed, if it's a named pipe, opening it
+     * for reading actually _causes inconvenience_ in its own right,
+     * even before the question of whether it gives misleading
+     * information.)
+     */
+    if (S_ISREG(st.st_mode) && st.st_size > 0) {
+        return TRUE;
+    }
+
+    return FALSE;
 }