Text complexity is the construct that readability formulas attempt to operationalise. It refers to the cumulative load a written text places on a reader, integrating surface-linguistic features (sentence length, lexical frequency), discourse-level features (cohesion, anaphoric distance, rhetorical structure), and reader-side features (topic familiarity, motivation, purpose). The Common Core State Standards in the U.S. consolidated the field's working definition into a three-factor model that has since travelled into ELT discussions.
Three components recur across the literature.
Quantitative dimensions. Surface features that can be computed automatically: average sentence length, syllable count, lexical frequency, type-token ratio. These are what classical readability indices measure. They predict comprehension reasonably well at the sentence level and poorly above it.
Qualitative dimensions. Features that resist automation: levels of meaning or purpose, structure (chronological versus non-linear), language conventionality and clarity, and demands on reader knowledge. These require human raters and are typically operationalised through structured rubrics.
Reader and task considerations. What the specific reader brings — proficiency, motivation, topic familiarity — and what the task asks them to do. The same passage is harder when the question targets inference and easier when it targets literal recall.
Text complexity is the joint distribution of all three. Reducing it to a single number, even a sophisticated one like a Coh-Metrix composite, loses information that matters for fair item design.
Two lessons recur in the testing literature.
First, calibration must triangulate. A passage's difficulty cannot be set by a quantitative index alone. Professional test development pairs computational indices with expert raters scoring qualitative dimensions and with empirical pre-testing on a representative candidate sample.
Second, text complexity and item difficulty are not the same construct. A simple passage with an inference question is harder than a complex passage with a literal one. Treating text complexity as a proxy for item difficulty produces calibration drift that only item analysis can catch, after the test has already shipped.
For AI-assisted item generation this means text-complexity validation must run against multiple metrics, not just one, and final calibration must include either human review or empirical pre-testing data before items enter the live bank.