<p>Some artificial intelligence systems attempt to implement counterfactual reasoning using computer simulations. But differences between the logics of counterfactuals and simulations suggest that the reliability of simulation as a means of counterfactual reasoning may be limited. On the usual understanding of counterfactuals, ‘If a vehicle had decelerated, snow would have been white’ is true, and the actual world factors into the evaluation of counterfactuals with true antecedents. But we would expect ‘In a simulation of a decelerating vehicle, snow would be white’ to be false, and for the actual world not to factor into our consideration of what happens in simulations with true stipulations. This paper formulates a logic to model simulation-based reasoning and identifies necessary conditions for the logics of counterfactuals and simulations to align. The investigations imply that simulation-based artificial systems can reliably implement counterfactual reasoning only insofar as it has an adequate notion of similarity between worlds, relevance between propositions, and compatibility between stipulations. Implications for the prospects of simulation-based artificial intelligence are discussed.</p>