This is called misalignment

We may some day find out that the smarter the model is, the hard is to align it properly