Stanford researchers found AI models gave more praise to Black students and less constructive criticism to minority groups, a new study says.