We consider prediction and uncertainty analysis for the 'history matching' problem in petroleum reservoir evaluation. The unknown reservoir properties are represented on a fine three dimensional lattice. A 'reservoir simulator', solving a set of partial differential equations, takes the reservoir properties as input and gives the production properties as output. The history matching problem is to infer the reservoir properties from an observed production history. To run the reservoir simulator on the lattice size of interest is a computer intensive task, and this severely limits the number of runs that can be made.We formulate the problem in a Bayesian setting and, following previous suggestions in the statistical literature, consider the reservoir simulator as an unknown function. We propose a new and more realistic prior formulation for this function, combining a (faster) version of the reservoir simulator on a coarse lattice with parameters correcting for the bias introduced by the coarser lattice. We simulate from the resulting posterior by Markov chain Monte Carlo (MCMC). We present a case study inspired by the Troll field in the North Sea. Convergence and mixing properties of the MCMC algorithm are good. The case study demonstrates how the observed production history provide information both about the reservoir properties and about the bias correcting parameters used in the prior specification.