The Flesch Reading Ease formula has been around since 1948. Until recently, it was mostly a concern for textbook publishers and UX copywriters. Now it's a signal that AI detectors actively monitor — and understanding exactly why tells you a lot about how detection systems are engineered.
## The Formula and What It Measures
Rudolf Flesch's readability score is computed from two variables: average sentence length in words, and average word length in syllables. The formula is:
*206.835 − 1.015 × (words ÷ sentences) − 84.6 × (syllables ÷ words)*
The output is a value between 0 and 100. Higher scores indicate simpler text; lower scores reflect denser, harder-to-process writing. A score of 70–80 is readable at a 7th-grade level. A score of 30–50 is college-level. Most newspapers target 60–70. Academic writing typically sits in the 20–50 range. Context determines what "good" actually looks like.
## Why LLM Output Clusters in a Predictable Range
Language models are trained to produce clear, coherent, well-organized prose — and that optimization has a measurable side effect. Unmodified GPT output almost always scores between 45 and 65 on the Flesch scale. It rarely drops below 30 (too dense for the model's coherence objectives) or rises above 75 (too casual for most training distributions).
That range isn't inherently suspicious. Plenty of human writers land there too. The problem is *consistency*. Human writing is structurally uneven: a long tangled sentence, then a short one, then a fragment. Paragraph-level Flesch scores bounce around naturally. AI output holds a narrow band across the entire document because the model is constantly balancing clarity and coherence throughout generation. That stability is one of the fingerprints that detectors are built to catch. For a deeper look at the full detection pipeline, see [how AI detectors work](/blog/how-ai-detectors-work-2026) at the implementation level.
## Burstiness: The Underlying Signal
Researchers use the term "burstiness" to describe sentence-length variation in text. Human writers produce bursts — a dense analytical passage, a short punchy conclusion, an occasional fragment. AI text exhibits low burstiness because the model generates sentences of roughly similar length throughout a document.
Flesch score is a useful proxy for this. A document where the score barely shifts from section to section is exhibiting machine-consistent behavior — which is one of the root causes behind [AI detection false positives](/blog/false-positives-ai-detection), where human-written text gets flagged simply because the author happened to write unusually uniformly.
Pairing a flat Flesch range with low perplexity scores produces a compound signature that detection systems can identify with high confidence. The score alone is weak evidence; the combination is strong.
## Target Scores and Practical Adjustments
For academic writing, a Flesch score between 40–60 is a reasonable baseline — complex enough to signal analytical rigor, accessible enough to remain readable. But the absolute value matters less than the *variance across sections*. A document that scores 55 uniformly from introduction to conclusion looks like machine output. One that swings between 40 and 70 looks like a person was thinking through it.
Several techniques shift the score organically:
- Alternate long analytical sentences with short conclusions. "The data suggests a strong correlation. That matters."
- Use contractions occasionally — "that's why," "here's the thing," "it doesn't mean" — they reduce the syllable-per-word average.
- Let some paragraphs run dense and technical. Let others breathe. Register shifts are a distinctly human signal.
- Open sentences with "But." or "And." periodically. It signals authorial voice rather than pattern completion.
- Dissolve some bullet lists back into prose instead of stacking five consecutive items.
Each of these structurally shifts the Flesch score across sections, which is the actual goal — not hitting a specific number, but producing measurable variation.
## Why Synonym Substitution Doesn't Work
Synonym swapping leaves Flesch scores unchanged because the metric depends on sentence structure and syllable counts, not semantic word choice. Replacing "utilize" with "use" moves the needle slightly, but swapping "demonstrate" for "show" does nothing to sentence length. Readability metrics operate at a structural level.
[WriteMask](/dashboard) addresses this by restructuring sentences, varying their length, and introducing the rhythmic variation that causes section-level Flesch scores to fluctuate the way human writing does. That structural-level rewriting is the reason it achieves a 93% pass rate — detectors measure deeper patterns than vocabulary, and the rewriting has to reach that layer to matter.
## Tooling for Inspection
To audit your own output, WriteMask's [readability checker](/readability) surfaces the Flesch score alongside other readability metrics in one view. Running text through the [free AI detector](/detect) gives a broader diagnostic — perplexity, burstiness, and structural consistency together, rather than relying on a single score in isolation. For students evaluating which approach fits their specific situation, [the guide on AI humanizers for students](/blog/best-ai-humanizer-for-students) maps out the tradeoffs by use case.
## The Core Takeaway
The Flesch Reading Ease score functions as both a direct and indirect input to detection systems. The issue isn't your absolute score — it's whether that score stays suspiciously flat across your document. Text that maintains a stable readability signature from start to finish exhibits the same structural characteristics as LLM output, regardless of who actually wrote it.
The fix is structural, not cosmetic. Write in a way where your Flesch score actually moves across sections — because that variation is what the math looks like when a human brain is driving the prose.
Originally published on WriteMask
Top comments (0)