We reanalyze existing paleodata of global mean surface temperature ΔTg and radiative forcing ΔR of CO2 and land ice albedo for the last 800,000 years to show that a state‐dependency in paleoclimate sensitivity S, as previously suggested, is only found if ΔTg is based on reconstructions, and not when ΔTg is based on model simulations. Furthermore, during times of decreasing obliquity (periods of land ice sheet growth and sea level fall) the multimillennial component of reconstructed ΔTg diverges from CO2, while in simulations both variables vary more synchronously, suggesting that the differences during these times are due to relatively low rates of simulated land ice growth and associated cooling. To produce a reconstruction‐based extrapolation of S for the future, we exclude intervals with strong ΔTg‐CO2 divergence and find that S is less state‐dependent, or even constant state‐independent), yielding a mean equilibrium warming of 2–4 K for a doubling of CO2.