Diverse methods have been developed to visualize the representations learnt by artificial neural networks. Such visualization methods however are not easily adapted to biological systems. This prevents direct comparisons between the internal representations of artificial and biological neural systems. One promising approach to visualize sensory representations in biological systems is reverse correlation. Behavioural reverse correlation experiments leverage pareidolia, our tendency to detect spurious signals in noise, such as seeing shapes, objects, or faces in clouds. In these experiments, observers report whether they detect a signal—e.g., the letter “s”—in noise. Just as some clouds happen to resemble known shapes, some noise samples resemble the signal an observer has in mind. Averaging over noise samples in which an observer spuriously detected a signal yields a “classification image”—a visualisation of their representation of the signal. Averaging noise however has several drawbacks: the method is slow to converge, produces blurred reconstructions, and cannot tease apart competing representations. We present a genetic algorithm approach that addresses these issues. We generate image populations by crossbreeding noise samples in which observers detect a signal. This approach converges faster and yields sharper reconstructions than reverse correlation, and is able to recover competing internal representations. Deep neural network image classifiers are able to correctly interpret the classification images generated by human observers, allowing us to “mind-read” which numerical digits observers are thinking of. This method could potentially recover—using equivalent stimuli and procedures—the internal representations of behaving organisms, neurons or neural populations, and units in neural network models, thus providing a powerful tool for comparing neural computations across biological and artificial brains.