AI detection is a classification problem. Detectors don't read your text — they compute statistics on it. Understanding which statistics they measure is the difference between effective rewriting and wasting an hour on edits that move your score by 11 points.
That gap matters more than most people realize. A **2023 Stanford University study** found that AI detectors incorrectly flagged non-native English speakers' writing as AI-generated at rates as high as **61.3%** — on text that was entirely human. Meanwhile, students who manually rewrite ChatGPT output often find it still gets caught. Both failure modes share the same root cause.
## The Two Metrics Detectors Actually Measure
Detection algorithms primarily operate on two signals: **perplexity** and **burstiness**. Perplexity measures how predictable each word choice is relative to what came before. Burstiness measures variance in sentence length across the document.
GPT-4 is optimized to select the statistically most probable next token at each step. The output is smooth, consistent, and metronomic — which means low perplexity and near-zero burstiness. Human writing is noisier. We make unexpected lexical choices. Sentence length lurches. We shift register mid-section. Detectors model that noise as a feature, not a bug. To get a deeper look at how these classifiers are built and where they break down, see [how AI detectors actually work](/blog/how-ai-detectors-work-2026).
## Why Synonym Replacement Doesn't Move the Needle
Most manual rewriting targets vocabulary — swap "utilize" for "use," restructure a phrase here and there. The problem is that synonym-level edits change surface tokens while leaving the underlying statistical distribution intact. The perplexity score barely shifts because you're still selecting from the same high-probability word neighborhoods.
Empirically: a 2024 analysis by Originality.ai ran 200 samples of manually edited AI text through GPTZero, Turnitin, and Copyleaks. Synonym-only editing reduced average detection scores by **11 percentage points**. Text starting at 90% AI-detected comes out at 79% — still flagged, with nothing to show for the editing time invested.
## Editing Techniques That Actually Affect Perplexity and Burstiness
The interventions that produce measurable score changes target the two signals directly:
- **Shatter compound sentences into fragments.** ChatGPT defaults to multi-clause balanced constructions. A two-word sentence after a long one generates a burstiness spike that reads as human. It also lands harder.
- **Inject specific, verifiable details.** Named examples, looked-up statistics, concrete references. These push word choices into lower-probability territory — higher perplexity — because specific nouns and numbers are inherently less predictable than generalizations.
- **Deliberately engineer sentence-length variance.** Short. Then something longer that takes its time working toward a conclusion before arriving. Then medium-length. Pattern this consciously rather than letting it happen by accident.
- **Shift formality levels mid-section.** AI maintains one consistent register throughout an entire document. Humans don't. Alternate between technical phrasing, direct plain language, and the occasional casual aside.
- **Delete the first sentence of each paragraph.** ChatGPT consistently opens with setup and throat-clearing. Strip it. The paragraph almost always reads better, and you've eliminated a predictable structural pattern.
Applied correctly across a 1,000-word document, these techniques require 30–45 minutes of focused editing — assuming you know precisely what to target, which most people don't. A step-by-step walkthrough of this process for a specific platform is in the guide on [how to humanize ChatGPT for Turnitin](/blog/humanize-chatgpt-for-turnitin).
## Algorithmic Rewriting: How WriteMask Approaches the Problem
[WriteMask](/dashboard) attacks perplexity and burstiness at the statistical level rather than the vocabulary level. Because the underlying signal distribution shifts — not just the word choices — the output reads naturally rather than like a thesaurus was run over it. The result is a **93% pass rate** across Turnitin, GPTZero, and Copyleaks.
The workflow is: paste your ChatGPT text, run humanize, then verify through the [free AI detector](/detect) before submitting. That final verification step isn't optional — every document is different, and you want the data before it matters. WriteMask preserves your argument structure and content; only the statistical fingerprint changes. Your ideas come back intact.
## When Your Own Writing Is the Problem
Sometimes the source isn't ChatGPT. Structured, clear writing — or writing produced by non-native English speakers — can trigger the same low-perplexity flag that catches AI output. This is the [AI detection false positive](/blog/false-positives-ai-detection) problem, and it affects a larger population than most detection discussions acknowledge, as the Stanford data makes clear.
If you're uncertain whether detection is triggering on AI text or on your own writing style, run a detector first. Diagnose before you edit. Applying AI-targeting rewrites to genuinely human text is the wrong fix for the wrong problem.
Detection evasion is an engineering problem with known variables: perplexity and burstiness. Synonym swaps don't touch those variables. Structural intervention does — whether you apply it manually with the techniques above, or use a tool built to target those metrics directly.
Originally published on WriteMask
Top comments (0)