They only ignore the first byte of an invalid byte sequence.

For example, b'\xff\n'.decode('gb2312', 'replace') gives '\ufffd\n' instead of
'\ufffd'.