Preferably the latter, but human communication is inherently ambiguous and heavily sensitive to context, and so actually implementing this is really hard* and t

Preferably the latter, but human communication is inherently ambiguous and heavily sensitive to context, and so actually implementing this is really hard* and thus the former is actually (slightly) feasible at the moment. *In fact, probably impossible: how often does one have to actually ask the author what they actually meant? Not appropriate for offline computer processing.