Non‐linear inversion of controlled source electromagnetic data is non‐unique. Inversion ambiguity and uncertainty grow with model complexity and limitations in sensitivity, for example when imaging deep targets. Uncertainties should be estimated. The best way to do that is by statistical inversion techniques. Because of the large number of parameters of 3D models, these methods are too costly for 3D applications. But neglecting or underestimating uncertainties often leads to erroneous interpretation and poor exploration decisions. The number of inversion parameters can be dramatically reduced while still being able to describe a complex 3D subsurface. To obtain that, I represent prior information by an input ensemble of geologically reasonable prior models. The model space is defined by the principal directions of this ensemble, where the inversion parameters are the coefficients of the different components. In this model space, inversion becomes much more efficient and can produce an updated ensemble of posterior models at the same computational cost as a controlled‐source electromagnetic inversion project using conventional non‐linear inversion techniques. I present examples where unique Gauss–Newton inversions lead to difficult interpretations and poor conclusions about hydrocarbon presence and saturation. The new approach provides a quantitative estimation of resistivity ambiguities and uncertainties leading to better interpretation and safer decisions. The examples use 2D synthetic controlled‐source electromagnetic data and 3D controlled‐source electromagnetic field data.