Here's a paper going into considerable detail about the new ML (not AI) model. It uses the ERA5 reanalysis charts as its training source and, on the specifics tested, does indeed provide higher accuracy than the op run. No comparison was made to the ensemble output, but it's interesting to see nonetheless.
The paragraph below really stood out to me - it seems that the ML model will attempt a blended solution if faced with two conflicting scenarios - and that's something we humans have often done! Most winters will see the GFS vs ECM, or similar, and it's often the case that somewhere between the two extremes is what verifies.
The non-linearity of weather dynamics means there is increasing uncertainty at longer lead times, which is not well-captured by a single deterministic forecast. ENS addresses this by generating multiple, stochastic forecasts, which approximate a predictive distribution over future weather, however generating multiple forecasts is expensive. By contrast, GraphCast’s MSE [Mean Squared Error] training objective encourages it to spatially blur its predictions in the presence of uncertainty, which may not be desirable for some applications where knowing tail, or joint, probabilities of events is important. Building probabilistic forecasts that model uncertainty more explicitly, along the lines of ensemble forecasts, is a crucial next step.
https://www.science.org/doi/10.1126/science.adi2336
The current model day 6 verification stats can be seen below: as is usually the case, ECM remains in the lead, with MetO a close second. It's interesting that improvements to the accuracy have levelled off over the past 10 years, which is no surprise - there are limits as to how far you can go with increasing resolution before it starts backfiring on you!
https://emc.ncep.noaa.gov/users/verification/global/gfs/ops/monthly_long_term/all_models/