Accessible text is important, for learners in school, and for adults in the world. A 2017 study reviewed the reading level of health-related wikipedia articles using a variety of measures and found they score around grade 14. Too high for a typical patient to understand.
The same problem is even more challenging and important in education. A complicated text may prevent a student from accessing the content they need to access for their learning.
Some have argued that AI could help to re-level these texts for any reading level. While this is probably true, it's a little harder than we might think at first. AI, is trained on readings at high reading levels, and as a result, tends to write at those levels. A 2025 study found reading grade levels of 11.40 for ChatGPT, 11.87 for Copilot, and 12.13 for Gemini.
But maybe if we prompt it to re-level it? Does AI really know what grade level means, such that if you ask it to write on a certain reading level, it can? We tested this and found that, on average, AI was about two grade levels too high, and at no point did it alert the user, saying "this may not be correct."
There's a litany of statistical methods for measuring the readability, complexity and "reading ease" of a text:
There are also proprietary measures, such as Renaissance's ATOS score, MetaMetrics' Lexile Levels, and our QuestionWell Readability Score.
These readability scores fall into a couple buckets.
FKGL estimates a grade level, and FRE produces a score normalized from 0-100. While the 0-100 range may be useful in some applications, it introduces confusion if you then convert that number into a grade level, which often does not match FKGL.
So, even on these straightforward, objective calculations, the mapping to grade level is still unclear.
The remainder of the readability measures rely on the familiarity of vocab instead of, or in addition to, text complexity measures. Dale-Chall is the simplest example, it considers two things:
While text complexity measures may include a definition of difficult words, like all ≥3 syllable words, these heuristics are limited. Many easy, grade appropriate words are 3 syllable, for example "animal". Dale-Chall improves on this by including a list of 3000 "familiar words" which should be appropriate for 4th graders. When applied to grades above or below fourth, and proportion of these familiar words is used to compute the score. In a 10th grade reading, there should be relatively fewer familiar words (for a 4th grader).
Building on the shortcomings of Dale-Chall, there are more sophisticated scores, such as ATOS, which use a larger word list which covers all grade levels. The Renaissance website explains how ATOS works at a high level (average word length, average word grade level, average sentence length, and book length), however, they do not make a wordlist or formula available. Lexile levels, developed by MetaMetrics, estimate text complexity primarily using word frequency and sentence length, with a more complex model for early-reader texts. The underlying formula and parameters needed to independently reproduce an official Lexile measure are proprietary.
Historically, we have used FKGL as our primary readability metric (but by no means our only objective). We are now introducing a QuestionWell Readability Score, which combines word age-of-acquisition with text complexity measures to estimate grade level.
Using a complete vocabulary list also gives us the ability to selectively preserve vocab words, by including them in the list of familiar words. This means you can differentiate an academic text while preserving important but challenging vocab words for the reader to learn.
1. Find a story about current events that's relevant to what you're teaching (e.g. https://www.nytimes.com/2026/08/06/science/ai-viruses-bacteria-arc.html)
2. Copy-paste the story into the QuestionWell New Quiz form, and select "Differentiate Reading"