The increase of CO₂-doubling-induced warming (climate sensitivity) in the latest Earth system models (ESMs) is primarily attributed to a larger extratropical cloud feedback. This is thought to be partly driven by a greater ratio of supercooled liquid-phase clouds relative to all clouds, termed liquid phase ratio (LPR). Yet global-scale multimodel evaluation of CMIP6 cloud phase representation is still missing. Here we first analyze the relationship between LPR and temperature in different types of observations (satellites, ground-based and in situ). We then evaluate for the first time global-scale LPR of ESMs from a fixed set of modeling centers across three coupled model intercomparison project (CMIP) generations (CMIP3, CMIP5 and CMIP6) against satellite observations. We find that LPR is greater in CMIP6 than in previous CMIP ESMs, and is overestimated rather than underestimated when compared to observations using an instrument simulator. Overall, models with prescribed temperature-dependent phase partitioning drive the multimodel bias while models with more complex microphysics fall within the observational range of uncertainty. Furthermore, the LPR overestimate is driven by an excess of liquid clouds in the extratropics and a lack of ice clouds in the tropics. Finally, in our multi-ESM-based analysis, greater LPR (i.e., more liquid clouds in ESMs) corresponds to stronger negative cloud feedback, in contradiction with single-ESM-based studies.