Not my observation, but I've seen chatter about how the AI ECMWF has been showing more run to run consistency (i.e. no wild swings).
Of course the question is whether this is a real confidence builder, or only perceived because of the run to run stability. In theory being able to quickly produce a large ensemble might be able to truly give us the kind of dispersion we need for probabilistic forecasting. BUT the problem with AI is that it needs history to train on, so if it has never happened before the model isn't going to create it like a deterministic might.
You aren't wrong about the day 4 or so difference though.