Here's a counterintuitive result that shows up repeatedly in AI content detection: the more deliberately polished someone's writing is, the more likely it is to get flagged as machine-generated. The culprit isn't word choice or topic — it's a 1948 readability formula that modern detectors have quietly co-opted as a statistical signal.
Understanding how this works requires a short detour into how AI detectors actually evaluate text — and why the metric they're exploiting was never designed to distinguish humans from machines.
The Flesch Reading Ease Formula
The reading ease scale — specifically the Flesch Reading Ease score — produces a value between 0 and 100 using exactly two inputs: average sentence length and average syllables per word. The score is inversely proportional to complexity. A score of 90–100 approximates children's book prose. A score of 60–70 is the target range for most web content. Academic and legal writing typically lands below 30.
Rudolf Flesch developed the formula in 1948 to help educators evaluate textbook accessibility. It's now embedded in Microsoft Word, the Hemingway App, and dozens of other writing tools. For most professional writers, a high Flesch score functions as a proxy for quality — clear, efficient, accessible. That assumption is where the problem starts.
What Detectors Actually Measure: Variance, Not Score
AI detectors are not reading your content semantically. They're running statistical analysis on patterns — and one of the most reliable patterns they exploit is readability variance across paragraphs. This is the core mechanism that explains how AI detectors work at an implementation level.
Language models, even without explicit instruction, optimize toward consistent output. Feed a model a prompt and it will produce text where paragraph-level readability scores cluster in a narrow band. There's no natural oscillation — no punchy two-sentence paragraph followed by a dense, clause-heavy one. The distribution is tight.
Human writers, by contrast, produce high variance almost incidentally. A thought that needs room gets room. A point that can land in five words gets five words. Fragments appear. Sentence lengths scatter. The variance isn't intentional — it's a byproduct of writing without optimizing for smoothness.
Low variance in readability scores across a document is a statistical fingerprint of AI generation. High, unpredictable variance is a fingerprint of a writer who wasn't thinking about readability at all.
Case Study: When Professional Polish Looks Like Machine Output
Alex Chen spent six years as a content marketing manager writing everything manually — no AI assistance, no generative shortcuts. In early 2025, a new enterprise client began running submitted articles through an AI content verifier before publication. Three of her first four submissions came back flagged at over 80% AI probability.
She ran her articles through a readability checker to audit the paragraph-level breakdown. Her average Flesch score was 81 — excellent by standard benchmarks. The problem was the distribution: every paragraph scored between 78 and 84. Near-zero variance across entire documents.
She had spent years training herself to write consistently clean, readable prose. In doing so, she had produced text that was statistically indistinguishable from well-prompted model output. This is precisely why AI detection false positives disproportionately affect experienced writers — polish looks like optimization, and optimization looks like a machine.
Diagnosing the Issue
The Flesch formula has a narrow measurement surface. It evaluates sentence length and syllable count. That's the complete input set. It says nothing about originality, argumentative structure, rhetorical rhythm, or voice. A one-word sentence like "Go." scores a perfect 100. A long, image-driven sentence with subordinate clauses might score in the 30s. Neither number encodes authorship.
What matters for detection isn't where your score lands — it's how much it moves between paragraphs. A document where every section scores 79–83 raises a flag. A document where scores range from 45 to 91 across sections reads as human-authored, regardless of the average.
The Fix: Three Implementation Changes
Alex adjusted her workflow over roughly two weeks and her pass rate moved from one in four to consistently above 90%.
- She stopped enforcing uniform clarity. Long sentences stayed long when the idea required it. Short ones didn't get padded. Contractions she had previously edited out were left in. Fragments that felt right structurally were allowed to stand.- She audited before submission, not after rejection. Using WriteMask's free AI detector, she could identify which sections were triggering flags and adjust before sending — rather than receiving an unexplained rejection with no diagnostic path.- For any AI-assisted drafts, she ran the full document through WriteMask. The tool goes beyond synonym substitution — it restructures sentence-level patterns and introduces the kind of readability variance that emerges naturally from human writing. The result aligned with WriteMask's documented 93% pass rate across major detection platforms.
Practical Takeaways
The reading ease scale is worth understanding not as something to game, but as a window into what detectors are actually measuring. They're doing pattern analysis on your prose structure, not evaluating your ideas.
Concretely: vary sentence length with intent. Let a complex idea sprawl when it needs to. Let a simple one be blunt. Run a paragraph-level audit using a readability checker and look at the distribution, not the average. If you're getting flagged despite writing original content, the techniques covered in how to humanize text for AI detectors generalize well beyond Turnitin to most major platforms.
Alex still writes manually. She just stopped optimizing every paragraph for maximum readability. Ironically, that regression toward inconsistency is exactly what makes the output read as human.
Originally published on WriteMask
Top comments (0)