I think this paper on Full Stack Alignment is very important in an age of AI.

Some of my friends were co-authors!

It points out that revealed preferences can't show the user's aspirations, since addiction and other manipulation can influence a user's revealed preferences.

What we "want to want" and what we want are different.

Anything that is a quantitative metric of desire will be thin and miss the deep nuance of real human aspirations and desire.

Luckily, LLMs allow qualitative insight at quantitative scale; for the first time, computers can understand what we want to want, and help us achieve it… if the system is aligned with the user's interest.