Does humanizing lower readability? Flesch scores and what the run reports

Two numbers. The Flesch Reading Ease score is 206.835 minus 1.015 times the average sentence length minus 84.6 times the average syllables per word, on a 100-point scale where higher is easier; the Flesch Kincaid Grade Level is 0.39 times the average sentence length plus 11.8 times the average syllables per word minus 15.59, giving a US school grade (Microsoft Support, readability statistics in Word, read 9 November 2026). Word's own documentation is the source here because it is the implementation most writers meet.

The grade-level formula comes from a 1975 US Navy study. Kincaid, Fishburne, Rogers and Chissom recalculated three readability formulas for Navy use from the test results of 531 enlisted personnel at four technical training schools on two bases, who read 18 passages from Rate Training Manuals, and published the result as Research Branch Report 8-75 (Kincaid et al. 1975, Derivation of New Readability Formulas, University of Central Florida STARS repository, read 9 November 2026). The formula was built to grade training manuals for sailors. It knows nothing about meaning, grammar or authorship.

Because the formula has two inputs, a single rewrite can push them against each other. A tool that breaks one 30-word sentence into two 15-word sentences lowers the grade level by 0.39 times 15, about 5.9 points, from the sentence term alone. The same tool, swapping "use" for "employ" and "help" for "assist" across the paragraph, raises syllables per word and pulls the grade back up. Which effect wins depends on the tool and the mode, and that is the measurement no vendor has published.

Nothing on any vendor page checked on 8 or 9 November 2026 answers this with a number. Undetectable AI's home page describes its humanizer as a tool to "score as human written on AI detectors and improve readability", read 9 November 2026, which is a direction claimed without a figure. Netus AI lists a Readability Checker among its tools, priced by analysis on its plans (Netus AI, pricing page, read 9 November 2026), which lets a user measure the change but does not report one. Grammarly, QuillBot, StealthGPT, Phrasly, HIX Bypass, WriteHuman, GPTHuman, Clever AI Humanizer, BypassGPT and Walter Writes AI state no readability figure on the pages read.

What the pages do describe are the operations. Walter Writes AI's home page says its tool "varies sentence length and tone" and does "structure-level rewriting"; Netus AI says its Bypasser "rewrites the whole passage"; Grammarly says its humanizer scans for "an overly formal tone or repetitive phrasing", all read 9 November 2026. Each of those operations moves one of the two Flesch inputs. Varying sentence length moves words per sentence in both directions within one text. Replacing a repetitive word with a less common one tends to add syllables, because the common word in English is usually the short one.

No, and the confusion costs people. Readability formulas count syllables and sentence length. Detectors score perplexity, which is how predictable each word is given the words before it. Liang and colleagues tested seven detectors on 91 human-written TOEFL essays in 2023 and found an average false positive rate of 61.22 percent, with the unanimously flagged essays showing significantly lower perplexity than the rest (Liang et al. 2023, Patterns, read 9 November 2026). When the authors prompted a chatbot to enrich the essays' vocabulary, the false positive rate fell to 11.77 percent.

Vocabulary enrichment is the operation that raises syllables per word. So the edit that lowered the detector reading in that study is the edit that, by the Flesch arithmetic, makes a text harder to read. A plain, short-worded paragraph can be both highly readable and highly predictable, which is why plain writing by second-language students was flagged. A text does not get flagged for being easy to read; it gets flagged for being expected, and the two overlap often enough that writers mistake one for the other.

Weber-Wulff and colleagues' 2023 test of 14 detectors adds the systemic point: no tool reached 80 percent accuracy, and tool paraphrase cut average accuracy to 26 percent (Weber-Wulff et al. 2023, International Journal for Educational Integrity, read 9 November 2026). The paper did not report readability. Its paraphrased documents were rewritten by a tool on its default mode, which is the setting the method on this site also runs first.

The method computes the Flesch Kincaid Grade Level of every input and every output and reports the change beside each tool, per genre and pooled, with no points attached. The weights go to meaning kept at 35 of 100, errors added at 20, detector pass at 30, and limits and access at 15, because a text that got simpler is not better for it and a text that got harder is not worse. A cover letter and a research abstract sit in the same input set and want opposite reading levels, so a scored readability change would reward the tool that happened to match the editors' taste.

The editors' reading rubric, which decides the meaning and error scores, is on the reading rubric page, and the readability column will appear with the first run on the runs page. No detector score should be the only evidence in a misconduct case, and no readability score is evidence of authorship in either direction. The rest of this series is in the guides hub.