We revisit the empirically observed positive correlation between DIF and difficulty studied by Freedle and commonly seen in tests of verbal proficiency when comparing populations of different mean latent proficiency levels. It is shown that a positive correlation between DIF and difficulty estimates is actually an expected result (absent any true DIF) in the presence of systematically negative ICC asymmetry. Under such conditions, conditional upon sum score, correct responses to easier items are indicative of higher latent proficiency than correct responses to more difficult items. Negative ICC asymmetry, which can occur as a result of disjunctively interacting latent subprocesses (including disjunctive combinations of proficiency-based guessing and problem-solving processes, for example), is suggested to be widely present among items used in many verbal tests due to their general lack of problem-solving complexity and the anticipated presence of proficiency-related guessing. When systematic, negative ICC asymmetries result in a shrinkage of units at the lower end of the IRT metric when traditional symmetric IRT models are applied. We demonstrate by simulation an ensuing DIF-difficulty correlation artifact in application of a model-based DIF approach, as well as for the Mantel-Haenszel and Standardization DIF detection methods. A sensitivity analysis illustration demonstrates how assuming a fixed level of negative asymmetry in the ICCs effectively makes an observed positive correlation between DIF and difficulty seen when fitting symmetric ICCs go away.