It's naive to think that models simply reflect their data. They reflect their tuning far more than their data. If they only reflected data, AI safety wouldn't be necessary.