How (not) to study deviance sensitivity and predictive coding in auditory cortex

• Deviance sensitivity and prediction errors are widely reported in brain responses • Common control sequences can misclassify deviance sensitivity in oddball paradigms • Adaptation and forward masking can mimic “deviant greater than control” effects • Valid tests require explicit mechanistic models, not single-condition comparisons Brain responses to sound depend strongly on recent sensory context, reflecting expectations generated at multiple time scales. A widely used paradigm for probing such context sensitivity is the auditory oddball sequence, in which rare “deviant” sounds evoke larger responses than frequent “standard” sounds, giving rise to mismatch negativity (MMN) and stimulus-specific adaptation (SSA). The prevailing interpretation of these effects invokes deviance sensitivity and prediction errors. To separate deviance sensitivity from ‘refractoriness’ – use-dependent adaptation that produces differences in the responses to common and rare sounds – studies in the field have relied on comparing deviant responses in oddball sequences to responses to the same stimulus embedded in “control” sequences, most commonly random multi-tone controls and ordered “cascade” controls, which do not have deviance. Here we argue that using these comparisons for inferences about deviance sensitivity is fraught with difficulties and requires much care. Indeed, using established models of auditory cortical dynamics, we show that deviant>control is not necessary, and may sometimes even not be sufficient, for deviance sensitivity. Adaptation-in-narrowly-tuned-modules models can yield deviant>control under some design choices that occur in the literature, while network models with firing rate adaptation and synaptic depression can produce deviant>cascade without any memory-based representation of “currently valid regularities” or computations of prediction errors. We conclude that demonstrating deviance sensitivity or prediction-error signaling cannot rely on the simple comparisons between responses in pairs of conditions, and requires explicit, falsifiable mechanistic modeling with richer experimental tests.