Methodology
AI detectors don't use magic. They measure specific statistical properties of text that differ between human writers and language models. This page explains exactly what we measure, why those measurements detect AI, and how our humanization pipeline eliminates those signals.
On internal test samples, the engine raises Human Score from a 51–57 baseline to 67–85 across writing styles (academic best at 85, blog 75–77, casual 67). External detector results vary by detector version — see per-style breakdown below.
Live demo
Paste any text below to run our 8-dimension analysis. The tool runs entirely in your browser — nothing is sent to a server at this stage. You'll see an AI score and a breakdown of which patterns were detected.
Scoring system
Our AI score is a weighted composite of eight independent measurements. Each dimension targets a distinct property of AI-generated text. The weights are calibrated against labeled corpora — texts we know to be AI-generated versus texts written by humans.
Human writers naturally mix very short and very long sentences. A paragraph might go: two words. Then a much longer sentence that develops an idea with subordinate clauses and qualifications. Then another short one. Language models produce sentences at a predictably similar length — usually 18–25 words each. We measure this as the coefficient of variation (CV%) of sentence lengths. CV% below 40% strongly signals AI authorship.
"Each dimension is assessed with precision. The scoring system evaluates multiple factors. Results are presented in a clear format."
"Short. Then a longer sentence that shows natural variation in rhythm and length. Then short again."
Language models are trained on internet text and develop strong statistical preferences for certain words. "Delve" appears in LLM output at roughly 100x the rate it appears in natural human writing. "Crucial," "straightforward," "tapestry," "foster," "leverage," "embark," "realm," and "utilize" follow the same pattern. We maintain a list of 60+ overused AI terms. High density of these terms is a strong detection signal.
"It is crucial to delve into this realm and leverage straightforward approaches to foster meaningful results."
"You need to dig into this area and use simple methods to get real results."
AI models overuse logical connectors — "furthermore," "moreover," "additionally," "in conclusion," "it is worth noting," "consequently." These words signal structured reasoning, which LLMs produce at a rate far above natural human writing. Humans connect ideas more implicitly, or simply start a new sentence without a connector. We count transition word frequency per 100 words. Above 5 per 100 words is a yellow flag.
"Furthermore, this approach is effective. Moreover, it has been tested extensively. Additionally, the results confirm the hypothesis."
"This approach works well. We have tested it extensively and the results back it up."
Passive voice is not an AI-specific pattern, but AI models overuse it relative to casual human writing. "The analysis was conducted" instead of "we ran the analysis." "Results were observed" instead of "we saw." Formal academic writing uses passive voice legitimately, but when passive constructions appear alongside other AI signals, the combination is diagnostic. We measure passive voice sentences as a percentage of total sentences.
"The experiment was conducted, results were analyzed, and conclusions were drawn by the research team."
"The research team ran the experiment, analyzed the results, and drew their conclusions."
This dimension works in tandem with burstiness. We specifically check whether any sentences under 8 words exist in the text. Human writers naturally produce very short sentences — fragments, punchy one-liners, abrupt statements for emphasis. LLMs rarely produce sentences this short unless specifically prompted. Complete absence of short sentences is a weak but consistent AI signal.
Em dashes are a stylistic marker that some models — particularly GPT-4 and Claude — overuse dramatically. A human writer might use one or two em dashes per page. AI-generated text sometimes places em dashes in nearly every paragraph. We flag texts with high em dash frequency relative to sentence count. This dimension matters specifically for GPT-4 and Claude output, which show this pattern most strongly.
LLMs default to structured output — numbered lists, bullet points, symmetrical three-part constructions. "First... Second... Third..." patterns. "Not only X, but also Y" constructions. Balanced parallel clauses. These structures signal organized, machine-generated reasoning. Human writing is messier — ideas trail off, points get developed unevenly, lists are less common. High parallel structure density is one of the strongest AI detection signals we measure.
"First, analyze the data. Second, identify the patterns. Third, draw your conclusions based on the evidence gathered."
"Look at the data, see what patterns show up, and go from there."
Contractions (it's, don't, you'll, we're) signal casual register. LLMs writing in formal mode avoid contractions entirely, producing text that reads as unnaturally stiff. Low contraction frequency combined with other signals suggests AI authorship. Note: this dimension is register-dependent — academic writing legitimately avoids contractions — so we weight it lower and primarily use it as a tiebreaker when other signals are mixed.
"It is important to understand that these patterns do not appear in isolation. They are part of a larger system."
"It's worth knowing these patterns don't appear alone. They're part of something bigger."
Key insight
Burstiness — the term researchers use for high variance in sentence length — is the single most reliable indicator of human authorship. The reason is structural: language models are trained to produce coherent, well-organized text. That training systematically pushes toward uniform sentence length. A human writer doesn't optimize for coherence at the sentence level — they write until the thought is expressed, which produces wildly variable lengths.
The coefficient of variation (CV%) measures this directly. CV% is the standard deviation of sentence lengths divided by the mean, expressed as a percentage. Human writing typically has CV% between 50–90%. AI-generated text clusters between 20–40%. This single measurement alone can separate human from AI-authored text with roughly 80% accuracy.
This is why our rewrite instruction attacks rhythm first: sentence length must be uneven and unpredictable, a long winding sentence next to a three-word one, never a run of same-shaped sentences. What we deliberately do NOT do anymore is enforce a CV% target with post-processing rules — engineered burstiness has its own detectable signature, which is exactly what our benchmark caught in the old pipeline.
Per-token probability
Academic AI detectors like Turnitin use a different approach from pattern-based tools: they measure per-token perplexity, which is a measure of how "surprising" each word choice is relative to what a language model would predict. LLMs tend to choose highly probable tokens — words that are the obvious continuation of the phrase. Human writers make unexpected word choices more often.
Low perplexity = each word was predictable = likely machine-generated. High perplexity = word choices were surprising = likely human. This is why simple synonym swapping doesn't work: if you replace "utilize" with "use," you've changed one high-probability word for another high-probability word — the perplexity stays low.
Refrazr addresses this by rewriting entire sentence structures. When the grammatical construction changes, the conditional probability of each subsequent word changes too. Structural rewriting shifts per-token statistics in a way that vocabulary substitution cannot. We make no promises about any specific detector's verdict — detectors retrain constantly — but this is the difference between rewriting and paraphrasing: one changes the writing, the other swaps words inside it.
Test results
Our benchmark (July 2026): 18 frozen AI-generated samples across six content categories — blog, landing, FAQ, academic, product, data-heavy — rewritten by the production pipeline and judged blind by Claude Opus on two questions: does this read as written by a person, and is every fact from the source intact. Genuinely human-written control texts score 9–10 with the same judge, which is how we know the judge itself is calibrated. The old rule-based engine and the current naturalness rewrite were scored on the identical corpus.
| Metric (Claude Opus judge, blind) | Old rule-based engine | Current naturalness rewrite |
|---|---|---|
| Outputs judged "written by a human" | ≈15–25% | 78% |
| Median human-likeness (0–10) | below bar | 7 |
| Meaning preserved ≥8/10 on every text | not gated | yes — hard gate |
A strict frontier-model judge is our quality bar, not a detector score. External detector results (GPTZero, Sapling, Turnitin, Originality.ai) vary by detector version, update frequently, and we make no claims about them. Individual texts vary — dense statistics and marketing copy are the hardest categories for any rewrite engine, ours included.
The pipeline
The eight dimensions above are our detection methodology. They're what the free AI detector on this site measures, and they're what taught us what makes text read as machine-written. The humanizer doesn't run through that checklist anymore, though. It used to: score the input, stuff the findings into the prompt, then sweep the output with a battery of cleanup rules afterward. In July 2026 we put that whole pipeline in front of a strict reading judge, and lost. Most of it got deleted, because the mechanical fixes we'd built were themselves the strongest AI signal in the text. What's left now is a lot simpler.
A fast heuristic pass sorts the draft into easy, medium, or hard to rewrite, based on question density, list structure, formality, and number density. No LLM call yet, nothing rewritten. That classification comes back as metadata alongside the result, and it decides the routing.
From there, the whole draft goes to the model in a single call, with one compact instruction. Write this the way a sharp person actually would: uneven sentence lengths, plain verbs, one clear voice. Underneath that sits one unbreakable rule: every name, number, date, and claim from the source stays exactly as written, and nothing new gets invented. No sentence quotas, no forced contractions, no word-count targets. For longer documents, the text gets split into large sections, so the model always has enough context to keep the argument in one piece.
After that, the output gets one light pass of cleanup: strip any leftover model preamble or meta-commentary, normalize the spacing, swap out stray em dashes. That's the whole post-processing layer now. No vocabulary swap list, no burstiness enforcement, no scoring-and-retry loop. Whatever the model writes is what you get.
We didn't simplify the pipeline to save money. We simplified it because the complicated version lost a fair test. We built a benchmark: real AI drafts, rewritten by each candidate pipeline, judged blind by a strict frontier model on exactly two questions. Does this read like a person wrote it? And is every fact from the source still intact? Our own four-stage pipeline, with its 150+ cleanup rules, kept failing that first question. Forced contractions, engineered sentence rhythm, vocabulary swaps, they all read as exactly what they are. A single honest rewrite instruction running on a cheaper model beat it on both counts. So we shipped the simple version and deleted the rules.
References
On false positives in non-native English writing — the strongest argument for humanization being a fairness tool, not a cheating tool: Liang et al. (2023), Stanford NLP, "GPT detectors are biased against non-native English writers" — found commercial detectors flagged TOEFL essays from non-native English speakers as AI-generated 61% of the time, while flagging native-speaker essays at near-zero rates. Available at arxiv.org/abs/2304.02819.
On the structural limits of perplexity-based detection — the research context for our deep-rewriting approach: Sadasivan et al. (2023), University of Maryland, "Can AI-Generated Text be Reliably Detected?" — argues that detector reliability breaks down against paraphrasing and structural transformations because they raise per-token entropy. Available at arxiv.org/abs/2303.11156.
On the limits of detector reliability in classroom contexts — relevant to the false-positive debate: Yale's Poorvu Center (2025) retracted multiple accusations against students after AI detector outputs were challenged. Coverage in The Yale Daily News, May 2025.
Internal corpus and raw per-sample scores: published in the project repo's internal test-results doc.
Try it
Free, unlimited browser analysis. No account needed to run the 8-dimension breakdown.
Humanize Your Text Free