Temporal confounds emulate multivariate fMRI measures of perceptual learning

Any study of learning must capture neural or behavioural changes over time. Stimuli are therefore presented more than once, before and after learning takes place-a sequence that cannot be randomised across time. This design creates vulnerability to temporal confounds that cause illusory neural effects. We investigate this issue in a classic 'one-shot' learning paradigm, in which ambiguous two-tone images become recognisable after viewing a photo cue. Across common fMRI designs, we found that spurious correlations driven by the temporal proximity of stimuli created strong artefacts in multivariate data. Among certain experimental designs, we found strong neural markers of perceptual learning depending on how stimuli were distributed across these artefacts. This paradigm-dependent effect was present not only in the primary visual cortex (V1), but also in a non-visual control region (S1), indicative of an artefactual origin rather than a genuine perceptual learning effect. We present a principled approach to correct for this temporal order confound. Once applied, the artefactual learning signature was eliminated from both V1 and S1, and while the corrected signal retained information about image identity in V1, this was absent in S1. Our work demonstrates how current multivariate neuroimaging approaches to understanding learning could be influenced by methodological artefacts that misdirect interpretations towards distributed neural processing, and offers a practical framework to distinguish true learning from measurement error.