The phenomenon of scalar diversity refers to the well-replicated finding that different scalar expressions give rise to scalar implicatures (SIs) at different rates. Previous work has shown that part of the scalar diversity effect can be explained by theoretically motivated factors. Although the effect has been established only in controlled experiments using manually constructed stimuli, there has been a tendency to assume that the marked differences in inference rates that have been observed reflect differences to be found in naturally occurring discourse. We explore whether this is the case by sampling actual language usage involving a wide range of scalar expressions. Adopting the approach in Degen (2015), we investigated the scalar diversity effect in a corpus of Twitter data we constructed. We find that the phenomenon of scalar diversity attenuates significantly when measured in a corpus-based paraphrase task. Although the degree of "scalar diversity" varies, we find that factors derived from theories of SI can explain nearly two-thirds of the variation. This remains the case whether the variation is observed in controlled experiments or in the context of natural language use. As for the remaining variation, we hypothesize that it may be due to a high level of uncertainty about whether adjectival scalar expressions should undergo scalar enrichment.