Gary Marcus

That large proportions of people have reading skills lower than high-end NLP a) has been demonstrated by the @oecd b) hasn’t been taken adequately into account by the #agi #singularity #Futureofwork panic crowds.

more generally @j2bryson what does your tweet even mean? is there a typo? what OECD study? what was the metric? what implications? having a hard time making sense of what you said.

The limit case is obvious – some people are fully illiterate so lots of AI gets more from text than they do. I'm illiterate in Chinese, for example. The OECD at their 2017 AI meeting talked about a study that showed that 70% of EU citizens can't do analyses top-end AI can 1/2

What my tweet means is the fact that AI can in some sense read more than humans (certainly books remember longer & better...) doesn't actually make humans redundant to humans. 3/2 #aiethics

reading is ultimately about extracting a cognitive model of something an author writes about. books can’t do that; current AI can’t do that either. for now, only (literate) people can do that.

I partially agree. Reading (or language understanding more generally) is about doing something sensible with linguistic input. AI can do that better than people in many circumstances. I agree books don't really do anything, and that humans are special to humans, but…

…but I don't believe that a "cognitive model" is a sensible delineation. I don't think there's one shape of model all humans & no machines use. This dichotomy is already harder than we generally acknowledge, but that this in no way undermines humans' special ethical status…