While I agree with most of the author's points, I'm not convinced at how possible it would be to have a product that "adopts the market-orientation and ad hoc n

While I agree with most of the author's points, I'm not convinced at how possible it would be to have a product that "adopts the market-orientation and ad hoc network of Twitter" in such a way that "harassment would be impossible." While deep learning and similar technologies can certainly make it much more difficult to harass someone, you can't rely on the content of someones message[0] nor can you rely on the network status of an individual, as spam accounts won't be readily differentiable from new accounts, and blocking new accounts from contacting people defeats the "ad hoc network" requirement. [0]: People are great at communicating meaning non-literally (see: definition of literally) and one of the ways that manifests itself is through slang, which is going to break anything but a so-close-it-might-as-well-be-AI system. If you ban a word or phrase, it's meaning will be communicated through another word or phrase. The insulting adversary has an advantage here, as they can use other real words and phrases that would otherwise be innocuous to insult you, it would be especially advantageous to use popular/common words/phrases as replacements (see: euphemisms). It's certainly possible to train a system in hindsight, but I highly doubt you could do so in real-time and any delay in reacting to new slang means that's when antagonists will strike. Hell, you could even do asynchronously defined insults, send many messages containing various words phrases and then announce, shortly after their delivery, what they mean (heh...).