A teacher pastes an essay into a detector and gets "87% AI-generated". An editor runs a freelancer's article through another one and gets "likely human". A job applicant is told their cover letter was flagged. In each case a single number is being asked to answer a question about a person: did they write this? That number cannot carry the weight, and the reasons are well documented.
What can be measured honestly is the text itself. Sentence length, vocabulary, hedging and punctuation are facts about writing, and they are the same facts that make a draft feel formulaic to a reader. This guide explains why authorship detection fails, what each measurable signal means, how the numbers are calculated, and how to use them to improve a draft rather than to accuse someone. The examples use the AI Writing Checker, one of the AI writing tools on Toolz, which reports six signals and deliberately gives no probability of AI authorship.
TL;DR: No tool can reliably tell whether AI wrote a text. OpenAI withdrew its own classifier in 2023 for low accuracy, and research shows detectors often flag non-native English writers as AI. What a checker can measure honestly is the writing: sentence-length variation (burstiness), average sentence length, overused vocabulary per 1,000 words, hedge phrases, em dash density and connective openers. The AI Writing Checker reports those six signals with the evidence behind each one, runs client-side, and is free, no signup.
Can an AI detector tell who wrote a text?
No. An AI detector is a statistical classifier that guesses at authorship from surface features of the text, and those guesses are wrong often enough that acting on them is unsafe.
The clearest evidence comes from the companies that built them. OpenAI released an AI text classifier in January 2023 and withdrew it in July 2023. The update on its announcement page says the classifier was no longer available "due to its low rate of accuracy". When the organization that trained the model cannot build a reliable detector for its output, a third-party score deserves little confidence.
The failures are not random, either. A 2023 study from Stanford researchers, GPT detectors are biased against non-native English writers (Liang et al.), ran seven widely used detectors over TOEFL essays written by non-native English speakers. More than half of those human-written essays were classified as AI-generated on average, while essays by US eighth-graders were classified far more accurately. The mechanism is simple: people writing in a second language tend to use a narrower vocabulary and more regular sentence structure, and regularity is exactly what detectors read as machine-like.
Two more problems make scores fragile. Light editing of a generated draft can move a score from "AI" to "human". And heavy editing of a human draft with grammar software can move it the other way. A measurement that flips under ordinary editing is not measuring authorship.
What can a writing checker measure honestly?
It can measure properties of the text that are true regardless of who produced it. The Toolz checker reports six:
| Signal | What it measures | How it is reported | Human-like end | Formulaic end |
|---|---|---|---|---|
| Sentence-length variation | How much sentence lengths differ (burstiness) | Coefficient of variation, as a percentage | 70% or more | 15% or less |
| Average sentence length | Mean words per sentence | Words | 14 or fewer | 30 or more |
| Overused vocabulary | About 40 words models overuse | Count and rate per 1,000 words | 0 per 1,000 | 20 or more per 1,000 |
| Hedge phrases | Phrases like "it is worth noting" | Count | 0 | 5 or more |
| Em dash density | Em and en dashes | Count and rate per 1,000 words | 0 per 1,000 | 10 or more per 1,000 |
| Connective openers | Sentences starting Moreover, Furthermore, However | Share of sentences | 0% | 30% or more |
Each signal is placed on a 0 to 1 scale between those two anchors, and colored as human-like (below 0.34), mixed (0.34 to 0.67) or formulaic (above 0.67). The anchors are editing thresholds, not statistical cutoffs for authorship. They mark where a reader starts to notice the pattern.
Every signal also lists its evidence: the actual sentence lengths, the words matched, the hedges found, the sentences that opened with a connective. If a number looks wrong, the evidence shows why.
What is burstiness, and why is it the strongest signal?
Burstiness is how much sentence length varies across a text. Human writing mixes short and long sentences. A three-word sentence lands a point. A thirty-word sentence walks through a qualification with its clauses. Generated drafts tend to settle into a uniform middle length, sentence after sentence of fifteen to twenty-two words.
The checker reports burstiness as the coefficient of variation: the standard deviation of sentence lengths divided by the mean. Dividing by the mean makes the figure independent of how long the sentences are overall, so a text of short sentences and a text of long ones can be compared fairly.
Here is the calculation for four sentences with lengths of 4, 17, 2 and 11 words:
mean = (4 + 17 + 2 + 11) / 4 = 8.5
squares = 4.5² + 8.5² + 6.5² + 2.5² = 20.25 + 72.25 + 42.25 + 6.25 = 141
variance = 141 / 4 = 35.25
std dev = √35.25 ≈ 5.94
CV = 5.94 / 8.5 ≈ 0.70 → "70% variation"
At 70% the text sits at the human-like end. Change those lengths to 15, 17, 16 and 18 and the standard deviation drops to about 1.12 on a mean of 16.5, a variation of roughly 7%, which is the formulaic end.
Burstiness is the strongest single signal because it is hard to fake without rewriting and easy to fix when you do: split one long sentence, merge two short ones, and add a very short sentence where a point deserves emphasis.
How are vocabulary, hedging, dashes and openers counted?
Overused vocabulary. The checker counts whole-word matches from a list of about 40 words that appear far more often in generated text than in human writing, including leverage, robust, crucial, comprehensive, holistic, meticulous, transformative and navigating. The count is converted to a rate per 1,000 words so short and long texts compare.
Hedge phrases. Ten patterns are counted, including "it is important to note", "it is worth noting", "it should be noted", "when it comes to", "at the end of the day", "needless to say" and "in today's fast-paced world".
Em dash density. Every em dash and en dash is counted and converted to a rate per 1,000 words. People typing on a keyboard rarely produce the character; models produce it constantly.
Connective openers. The checker counts sentences that begin with Moreover, Furthermore, Additionally, Notably, Importantly, Consequently or However, as a share of all sentences.
Average sentence length. The mean word count per sentence. Around 15 to 20 words reads comfortably; past 25 the reader starts working.
Sentences are split after a period, question mark or exclamation mark followed by whitespace, and words are runs of letters and apostrophes. Abbreviations like "e.g." can therefore split a sentence early, which slightly changes the length figures. The evidence list shows the lengths so you can spot it.
A worked example with every number
Paste this short text:
The report is ready. It covers three regions and explains why sales in the north fell by a fifth last spring. Read it. Moreover, it is important to note that the landscape is changing.
The checker reports 34 words and 4 sentences, with these signals:
| Signal | Value | Verdict |
|---|---|---|
| Sentence-length variation | 70% variation (lengths 4, 17, 2, 11) | Human-like |
| Average sentence length | 8.5 words | Human-like |
| Overused vocabulary | 1 instance (29.4 per 1,000 words): landscape | Formulaic |
| Hedge phrases | 1 found: "it is important to note" | Human-like |
| Em dash density | 0 dashes (0.0 per 1,000 words) | Human-like |
| Connective openers | 1 of 4 sentences | Formulaic |
Two lessons come out of this. First, rates on very short texts swing hard: one word in 34 is a rate of 29.4 per 1,000, off the scale. Paste at least a few hundred words before reading much into the vocabulary and dash rates. Second, the signals point at a specific sentence. The last one carries an opener, a hedge and a listed word, and rewriting it to "The market is changing." clears all three.
The overall label for this text is "Reads naturally", because the mean of the six scores lands just under the mixed threshold. That borderline result is a good illustration of why the overall label is a summary of the signals and not a grade.
How should you use a writing checker, and how should you not?
Use it on your own drafts, to find which paragraphs read as formulaic, and to learn what the patterns look like. Use it when editing contributed work, as a pointer to where a light edit would help.
Do not use it to decide whether a student cheated, whether a freelancer used a model, or whether an applicant wrote their own letter. The page says this in a warning box above the results, and the reason bears repeating: no detector can establish authorship, false positives fall hardest on second-language writers, and an accusation built on a score can do real harm to someone who did nothing wrong. If authorship genuinely matters, talk to the writer, look at drafts and revision history, or ask them to explain their reasoning.
There is also no score to aim for. If variation is low, vary your sentences. If hedging is high, cut hedges. Both improve the writing whoever wrote it.
How do you fix what the checker finds?
Each signal has a direct fix.
| Finding | Fix |
|---|---|
| Low sentence-length variation | Split one long sentence; merge two short ones; add a very short sentence for emphasis |
| High average sentence length | Break sentences at "and", "which" and "because" where the clauses stand alone |
| High vocabulary rate | Swap each listed word for a plain one: use, strong, key, complete |
| Hedge phrases | Delete the phrase; the sentence means the same thing |
| High em dash rate | Replace dashes with full stops, commas or brackets, whichever the sentence wants |
| Connective openers | Delete the opener; the order of the sentences already shows the connection |
Several of these can be applied automatically. The AI Text Humanizer removes hedges and openers, swaps overused words and replaces dashes, and lists each change so you can refuse the ones that do not fit. It does not change sentence length, so burstiness is still a manual edit. A practical loop is: check, humanize, edit sentence rhythm by hand, then check again.
If you want to see how much your edits changed, paste the before and after into a text diff checker. For length targets on a specific platform, the guide to counting words and reading time covers how word counts are calculated and why tools disagree.
Pattern checker vs AI detector vs readability score
These three tools answer different questions and are often confused.
| Tool type | Question it answers | Output | Reliable for |
|---|---|---|---|
| AI detector | Did a model write this? | A probability or verdict | Nothing consequential; accuracy is too low |
| Pattern checker | Which formulaic habits does this text have? | Measured signals with evidence | Editing and self-review |
| Readability score (Flesch and similar) | How hard is this to read? | A grade level or ease score | Matching text to an audience |
A readability formula and a pattern checker overlap on sentence length but look at different things beyond it. Flesch scoring counts syllables per word; the pattern checker counts specific words and constructions. Used together, they tell you whether a draft is hard to read and whether it reads as templated.
Is the text uploaded or stored?
No. Every measurement runs in JavaScript in your browser. The text is not transmitted, logged or stored, and the page works offline once loaded. When the text box loses focus, the page records a usage event containing the word count only, never the text.
That matters for the exact documents people most want to check: unpublished drafts, student work, manuscripts and client deliverables. The word lists and thresholds are fixed in the page, so the same text always produces the same numbers, which makes results reproducible if two editors compare notes.
To check a draft now, paste it into the writing pattern checker, read the evidence under each signal, and fix the two or three that matter most.
