Hah negative votes on this, so to expand a bit: Most HN responses are thoughtful and reflect some background. Timezones vary, but are self-consistent. Similarl

Hah negative votes on this, so to expand a bit: Most HN responses are thoughtful and reflect some background. Timezones vary, but are self-consistent. Similarly, different people gravitate to different topics. But when a reply is on a talking point associated with state misinfo and written in a misinfo/troll style (short, lies, unfalsifiable, red herring, pushes asymetric work to the other person, ...), I check to see if from an account mostly only making non-technical comments, time zone range, and other surprises. Unfortunately, no public way to do IP/browser-level correlation for account reuse and other common forensics. While such an account is unlikely to be a programmed bot, human troll farm employees are paid to do exactly this sort of campaign. NLP is great for catching funny behaviors and talking points, but without the forensic info, generally too hard to increase confidence. NLP on a message is like an alert, and NLP on the message history can tell you if someone acts like a troll, but metadata forensics are still typically needed to get proof points for saying it's a troll-farm-style troll. (Then again, if it walks like a duck and quacks like a duck, it's enough of a duck for many community leaders to treat it like a duck.) If HN's trust and safety team prioritized this stuff, they should have the needed metadata already, and the forensic analytics patterns are pretty established nowadays - you don't need to pay FB's $1B/yr on it. I'm sure they do some of it already in addition to the human moderation. Project Domino is all about building open AI so others can do the same without the high budget.