The problem of reasonably determining what someone is interested in from what they've previously read, skimmed and skipped shouldn't be "AI complete."
It only involves matching patterns to key-words. It might not be easy and the results might not be perfect but it seems possible. Using data from friends and finding the best correlates could help too. You could use the Netflix algorithm or other magic.
I suspect that's why today's readers don't satisfy - they aren't there yet and it's easy to imagine they could be.
I agree with your claim that recommendation algorithms aren't AI complete, but for different reasons.
"It only involves matching patterns to key-words." False. Recommendation might be far trickier.
Nonetheless, if you can do recommendation perfectly, it still doesn't mean that you can solve NLP, machine vision, control (robotics), planning, or any of the other major AI tasks.
But the point is that we have many examples of recommendation systems that do work reasonably well and working reasonably well is what I suspect would satisfy most users.
Further, a reader which could learn reasonably well from passive observation would give users the base do explicit tagging and rating.
It only involves matching patterns to key-words. It might not be easy and the results might not be perfect but it seems possible. Using data from friends and finding the best correlates could help too. You could use the Netflix algorithm or other magic.
I suspect that's why today's readers don't satisfy - they aren't there yet and it's easy to imagine they could be.