“A recent DeepMind paper demonstrated a method of collecting preferences from different demographic groups, training a language model on them, and aggregating the results using social welfare functions.”