Do goldfish like to be informed?

Most mammalian and avian species tested so far, including humans, prefer foretold over unsignalled future events, even if the information is costly and confers no direct benefit, a phenomenon that has been called paradoxical, or suboptimal choice. It is unclear whether this is an epiphenomenon of taxonomically widespread mechanisms of reinforcement learning, or if information-seeking is a dedicated cognitive trait, perhaps a precursor of human curiosity. We investigate whether a teleost fish that shares basic reinforcement learning mechanisms with birds and mammals also presents such preference, with the aim of dissociating food-reinforced learning from information-seeking. Goldfish chose between two alternatives, both yielding a 50% chance of reward 5s after being chosen. The ‘informative’ alternative caused immediate onset of either of two stimuli (S+ or S-) correlated with the trial’s forthcoming outcome (reward/no reward). Choosing the ‘non-informative’ option, instead triggered either of two uncorrelated stimuli (N1 or N2). Goldfish learned to discriminate between the different contingencies, but did not develop preference for the informative option. This shows that conditioning learning is not always sufficient, and the difference with birds and mammals supports the hypothesis that information-seeking, rather than simple conditioning, causes the paradoxical preference for unusable information shown by the latter.