There's absolutely a huge gap between user preference and model performanc that is widening by the minute. The more performant these models get, the more individual and syntactical preferences prevail.