Cloud-to-ground (CG) lightning poses significant risks to infrastructure, people, and ecosystems. However, while many studies have projected changes in total lightning under warming, future trends in CG lightning remain poorly constrained. Here, we combine ground-based CG lightning observations with six daily surface meteorological variables (mean temperature, total precipitation, short-wave radiation, relative humidity, pressure, wind speed) to predict CG lightning density over land at 0.5° resolution using an XGBoost machine learning model. The model accurately captures contemporary spatial and temporal lightning patterns and outperforms an alternative multiple linear regression model as well as three existing parameterizations based on convective precipitation or convective available potential energy times precipitation. Applying the XGBoost model to historical climate shows a 11% increase in global CG lightning frequency over the 20th century. Future projections using outputs from two climate models (MPI-ESM1-2-HR and UKESM1-0-LL) reveal large increases of 76-140% under a high-emission scenario (SSP3-7.0), equivalent to roughly 30% more CG lightning per degree of global warming. These increases are substantially higher than those projected by the linear model and previous studies. Absolute increases are largest in the tropics, while relative increases are largest at high latitudes, where lightning density often more than doubles. The largest uncertainties emerge in the tropics for UKESM1-0-LL climate, particularly in the Amazon and New Guinea, where future climate differs strongly from current conditions. Our results suggest that climate change may substantially amplify CG lightning frequency, underscoring the need for improved risk assessments.