Hacker News Reader

"Our ML system registered a hit on your account, which is almost always associated with policy violations, but we don't know what the trigger was, or why that set of data about your account is almost always associated with policy violations... it just is."

View on HN

"and we know the ML system always works, because those that get banned from it never get their account reinstated"

"and we know that it doesn't always work and don't care because despite our ML model being opaque and hard to introspect when we tested it against our database of known bad actors it performed better than both human review and our old heuristic based system."

There is an opposite side to this. If you have a human appeal process your accuracy drops to min(ml_model, humans) and I have bad news about which one of these is smaller.

We're on the cusp of independent artificial intelligence and it's incredibly lame.

Reply on news.ycombinator.com