Yeah hopefully! The only fear I’d have is you could end up in scenarios where what’s useful for humans will represent extra context that leads a model in an undesirable direction — but it’s worth testing in case I’m wrong!