Using a standard prediction of action task, we find that normally developing 3-yearold children successfully take into account a protagonist's false belief when asked a ‘look first’ question. When asked this same question in a true belief scenario, 3-yearolds also correctly predict a protagonist's action even though in this case the correct answer is the second (full) location rather than the first (empty) location. This rules out the use of a low-level response strategy. In a second experiment, children aged 3.5 years who failed a standard ‘think’ question passed a ‘look first’ question. A control group of older children with autism, who performed at the 3.5-year-old level on the ‘think’ question, showed no improvement in performance on the ‘look first’ question. Taken together, the two studies confirm that a minimal modification to the standard false belief task helps normally developing preschoolers to calculate the content of a false belief. Current neuropsychological models of ‘theory of mind’ development typically contrast autistic children with normally developing 4-year-olds who pass standard false belief tasks. Our present results extend current models by comparing autistic children with 3-year-olds who also fail standard tasks. Apparently, the two groups do not fail for the same reasons. Whereas 3-year-olds’ difficulties on theory of mind tasks appear to be due to performance factors, autistic children's difficulties appear to be caused by a deeper metarepresentational impairment.