I think that argument ignores that neural nets can learn the models that Pearl believes are important. And it's quite plausible that human reasoning works in similar way.
Pearl's career is all about causal networks, so he's slightly biased when it comes to survey like this.
