Adam Subel and Laure Zanna present new work assessing whether AI ocean emulators can reliably reproduce the ocean’s long-term response to shifts in climate forcings. Using the midHolocene experiment as a controlled testbed, they show that autoregressive, full-depth ocean emulators can generalize to new orbital forcings and capture the large-scale spatial response in the upper ocean, seasonal changes, and shifts in ocean variability, although they tend to underestimate their amplitude. The results also reveal important limitations: emulators struggle to reproduce the slow, internally driven evolution of the deep ocean, and strong performance on standard training metrics does not necessarily translate into a skillful response to changing forcings. Together, this work establishes the midHolocene as a valuable, ground-truthed framework for diagnosing and improving climate emulator performance before applying these models to more unfamiliar climates.