Comment by energy123

That's true, and interpretability is helpful for AI safety.