I think the core problem is having one universal metric. Take Uber or Lyft as an example. Some people like silence, some like conversation, some like music. If the ML models were better or if it were based on a web of trust we wouldn't need to have a score that treats a 4.5 as universally better than a 4.6 or similar.