Speakers tend to use reduced pronunciation, e.g. shorter duration, when words are previously mentioned, or predictable in context. Existing accounts of this phenomenon underspecify whether both giveness and predictability make independent contributions, and say little about the underlying cognitive mechanism. I propose and test the Activation Reduction Hypothesis (ARH), which states that any stimulus that activates representations used for language production should elicit reduced pronunciations. This unites givenness and predictability in a single plausible psychological mechanism, and makes novel predictions, which I tested in three experiments. The first experiment shows that linguistic stimuli elicit more reduction than non-linguistic stimuli, which also elicit reduction. The second shows that linguistic stimuli elicit reduction in the absence of strong predictability, suggesting a role for sheer activation. The third attempts to isolate this reduction at the conceptual level of representation, but shows little supporting evidence.