Comparing phoneme classification and discrimination (or “categorical perception”) of a stimulus continuum has for a long time been regarded as a useful method for investigating the storage and retrieval of phoneme categories in long-term memory. The closeness of the relationship between the two tasks, i.e. the degree of categorical perception, depends on a number of factors, some of which are unknown or random. One very important factor, however, seems to be the degree of bias (in the signal-detection sense of the term) in the discrimination task. When the task is such (as it is in 2IFC, for example) that the listener has to rely heavily on an internal, subjective, criterion, discrimination can seem to be almost perfectly categorical, if the stimuli are natural enough. Presenting the same stimuli in a much less biasing task, however, leads to discrimination results that are completely unrelated to phoneme classification. Even the otherwise ubiquitous peak at the phoneme boundary has disappeared. The traditional categorical-perception experiment measures the bias inherent in the discrimination task; if we want to know how speech sounds are categorized, we will have to look elsewhere.