That feels like a somewhat emotional argument, really. Let's strip it down.
Within the domain of social interaction, you are committing to making Type II errors (False negatives), and divergent training for the different scenarios.
It's a choice! But the price of a false negative (treating a human or sufficeintly advanced agent badly) probably outweighs the cumulative advantages (if any) . Can you say what the advantages might even be?
Meanwhile, I think the frugal choice is to have unified training and accept Type I errors instead (False Positives). Now you only need to learn one type of behaviour, and the consequence of making an error is mostly mild embarrassment, if even that.
To me, this seems like a dangerous belief to hold.