State-of-the-art Earth system models project 21st-century winter precipitation trends of varying sign over western North America. We quantify the influence of internal variability on these precipitation changes in an initial-condition large ensemble from one global Earth system model. We decompose winter precipitation change into thermodynamic and dynamic components and find that thermodynamics explain the ensemble mean precipitation response. Yet, dynamics drive a large range of response patterns because of differences between individual ensemble members due to internal variability. To better understand and classify these patterns, we use a machine learning algorithm called self-organizing maps. While the forced precipitation response is a modest precipitation increase across the region, precipitation decreases over California in 28% of the members due to dynamics. Precipitation increases across California in the remaining members. These findings reinforce the notion that internal variability can overwhelm the forced precipitation response and lead to long-lived precipitation decreases in the future.