By Stephen Milner · UtilityForge · Last reviewed: May 2026
The AI Writing Pattern Analyzer is a free, browser-based tool that checks any block of text for four statistical markers consistently found in AI-generated writing. It returns a heuristic likelihood score from 0% to 100%, a per-sentence heat map, and individual breakdowns for each marker. All processing runs locally in your browser. Your text is never transmitted to any server.
AI language models tend to produce sentences of similar length. Human writers naturally vary sentence length, mixing short, punchy sentences with longer, more complex constructions. This tool measures the coefficient of variation (CV): the standard deviation of sentence word-counts divided by their mean. A CV below 0.3 is a strong AI signal; human-written prose typically registers between 0.4 and 0.8.
Type-token ratio (TTR) is the number of unique words divided by the total word count. AI models trained on large corpora tend to reach for the same vocabulary repeatedly, producing a lower TTR than skilled human writers at equivalent text lengths. A TTR below 35% is consistent with AI output; above 55% is more typical of human writing.
The burstiness parameter B (adapted from Goh and Barabasi, 2008) measures whether sentence lengths follow a regular rhythm or an irregular, clustered pattern. B ranges from -1 (perfectly regular) to +1 (maximally bursty). Human writers naturally vary their pacing: short sentences for impact, longer ones for explanation. AI models produce more metronomic prose with B values close to or below zero. Values below -0.1 are flagged as an AI signal.
Certain phrases appear disproportionately in AI output. Common examples include "it is worth noting," "furthermore," "in conclusion," "delve into," "it is crucial to," and "that being said." The tool checks against a curated list of over 50 documented AI phrases and scores their density per 100 words.
| Score | Verdict |
|---|---|
| 0-29% | Likely human-written |
| 30-54% | Inconclusive |
| 55-74% | Possibly AI-generated |
| 75-100% | Likely AI-generated |
The overall score is a weighted average: sentence burstiness carries 30% weight because it is the hardest marker for prompt engineering alone to circumvent. The remaining three markers each contribute 25% or 20%.
These thresholds are heuristic. A 40% score does not mean 40% of the text is AI-generated. It means the statistical patterns fall between typical human and typical AI writing profiles.
Educators use it as a first-pass screen for student assignments before deciding whether to investigate further. Editors and publishers flag submissions that may warrant closer scrutiny. Content marketers audit AI-assisted drafts to ensure the final copy does not carry obvious AI fingerprints before publication. HR teams review writing samples submitted with job applications.
No statistical detector is definitive. Factors that can produce false positives include:
Conversely, heavily edited AI output or carefully crafted prompts can reduce scores significantly. Treat any result as a signal that warrants further scrutiny, not as proof of AI authorship.
All analysis happens in your browser using JavaScript. Your text is never uploaded to any server, stored in any database, or shared with any third party. There are no accounts, no tracking cookies, and no data retention.
The tool measures four statistical properties of your pasted text: sentence length uniformity (coefficient of variation), vocabulary diversity (type-token ratio), sentence burstiness, and the density of documented AI filler phrases. Each property is scored 0-100 based on how closely it matches typical AI writing profiles, and a weighted average produces the overall likelihood score. Processing runs entirely in your browser. Nothing is sent to a server.
No. The score is a heuristic indicator, not a verdict. Statistical patterns can appear in human writing, particularly academic or formal writing, and can be absent in heavily edited AI text. Use the result as a starting point for further evaluation, not as standalone evidence.
The markers are calibrated against output from GPT-class and Claude-class models. All current large language models trained on similar data tend to produce comparable statistical patterns, so the tool generalises well across AI writing assistants. It is not specific to any single model.
CV is the standard deviation of sentence word-counts divided by their mean. A value near zero means all sentences are roughly the same length, which is consistent with AI generation. Human writers produce more varied sentence lengths, typically yielding CV values between 0.4 and 0.8.
Burstiness (parameter B) measures whether sentences cluster in irregular patterns of length. Human writing is naturally bursty: writers group short sentences together for impact, then shift to longer, explanatory passages. AI models produce more evenly paced output with B values close to or below -0.1. The parameter is adapted from Goh and Barabasi (2008), who used it to characterise event timing in natural and social systems.
Common reasons include: the text is short (under 150 words); the writing style is deliberately formal or structured; or the same phrases recur across multiple sentences. Review the per-sentence heat map to identify which specific sentences are triggering signals, and check whether any matched phrases apply to your writing style.
A minimum of 50 words is required to run the analysis. For reliable results, 150-500 words is the ideal range. Very long documents may produce slightly skewed TTR scores, since unique-word density naturally decreases as any single text grows longer.
No. The entire analysis runs locally in your browser using JavaScript. Your text is never uploaded, logged, stored in a database, or transmitted to any external server or third party.