Reliable projections of wildfire are crucial for the development of efficient and effective adaptation and mitigation strategies. The lack of or limited observational constraints for Earth system model (ESM) outputs impairs the credibility of wildfire projections. Here, we present a combined machine learning and emergent constraint framework to constrain the future fire carbon emissions simulated by ESMs from the Coupled Model Intercomparison Project phase 6, using historical, observed joint states of fire-relevant variables. The new approach leads to improved representation of both the magnitude and spatial distribution of global fire carbon emission during the validation period. The observation-constrained ensemble projects a 4.1% (2.6%–7.2%) decade^-1 increase in the global fire carbon emission during the twenty-first century, to a lesser extent than the 6.0% (0.6%–9.4%) decade⁻¹ increase as indicated by the default ensemble. Moreover, the observation-constrained ensemble indicates a further enhancement of wildfire carbon emission in the historically fire-prone subtropical savannahs and tropical forests, and in the northeast United States and the Appalachian Mountains. Our new constraining framework provides an encouraging approach for correcting ESM biases and can be expanded to estimate reasonable evolution of other global and regional climate or ecosystem properties, especially those with extensive local impacts.