An AI writing tells checker is a style-hygiene scanner, not a detector. Paste a draft and it counts the habits that make prose feel machine-assembled — stock phrases, uniform sentence rhythm, hedge stacking, rule-of-three overuse, repeated openers and passive voice — then hands you a polish score and a prioritized fix list. It runs entirely in your browser, and nothing you paste leaves your device.
AI detectors are unreliable in both directions: fully human writing gets flagged, machine-written text gets waved through, and students have paid real prices for both errors. This page will not guess who wrote your draft, and you should not treat its score as one.
What it does instead is an editing pass. It finds the formulaic habits that make prose feel assembled rather than written — the texture a grader's eye catches after a stack of essays. The best fix is never cosmetic. It is adding your own examples, your data and your stance, so the sentences could only have come from you.
And if your course bans AI assistance, no scanner repairs that problem. Write the draft yourself, then use this page to tighten it.
Run the scan on your draft
Paste the draft below — 200 words or more gives every check enough to work with, and up to 20,000 words is accepted. Hit Scan draft and the six reports fill in under the score. If you just want to see how the tool behaves, load the demo draft: an invented, illustrative 450-word essay deliberately written to exhibit the tells. Your text is saved in this browser only (key csw-ais-v1), never uploaded, and Clear wipes it.
AI writing tells scanner
100% private — runs in your browser0 words — 200 or more gives every check enough to work with.
—
Fix these first
Uniformity ratio = standard deviation ÷ mean sentence length. Below 0.45 the draft reads as a metronome; the classic repair is to follow a 30-word sentence with an 8-word one.
Counted hedges include arguably, somewhat, rather (not "rather than"), quite, to some extent, it could be argued, generally, typically, perhaps, possibly, tend to, relatively and to a certain extent.
A rough heuristic: is / are / was / were / be / been / being followed by an "ed" word. It overcounts adjectives ("a marked increase") and misses irregulars ("was written" is caught, "was driven" too, "was put" is not). Treat it as a smell test, not a grammar ruling.
The report is text you can paste into your notes. It names counts and sentences, not judgments about authorship — the editing decisions stay yours.
One pass through the fix list usually takes ten to twenty minutes on a 1,500-word essay. The order matters: phrases first (cheap, visible), rhythm second (harder, more valuable), voice last (the part no tool can do for you). The sections below explain each tell, show before-and-after pairs, and end with the limits you should know before trusting any scanner — this one included.
What professors actually notice
Graders rarely open a detector first. They open your essay after thirty others, and pattern recognition does the rest. What raises an eyebrow is not one suspect word — it is sameness stacked on sameness: paragraphs that all run five sentences, claims that all retreat into hedges, examples that could belong to any essay on the topic because they name no course, no reading, no number. A draft can be grammatically clean and still feel assembled, the way a room can be tidy and still look staged.
Here is the honest complication: every tell on this page appears in fully human writing too. Students genuinely write "moreover" and mean it. Real conclusions sometimes contain three-part lists because the material really has three parts. That is exactly why the tell count matters more than any single flag, and why this page reports densities and ratios instead of a single verdict. The six checks, and what each one counts:
| Check | What it counts | What a fix looks like |
|---|---|---|
| Stock phrases | Hits on a curated list of about 60 formulaic phrases, per phrase, with the sentences that contain them. | Rewrite the claim around a specific fact — the phrase usually sits where a fact should be. |
| Rhythm | Mean sentence length, standard deviation, and the uniformity ratio between them. | Mix lengths deliberately: a 30-word sentence earns an 8-word follower. |
| Hedge stacking | Hedging words and phrases per 100 words. | One hedge per genuinely uncertain claim; cut the reflex ones. |
| Rule of three | "X, Y, and Z" triads per 100 words. | Keep triads that carry three real items; let two-item lists be two items. |
| Repeated openers | How often the same first two words start a sentence. | Vary the entry: start from the subject, the data, or the objection. |
| Passive voice | A heuristic estimate per 100 words, with the matched snippets. | Name the actor — except in Methods sections, where passive is standard. |
Notice what the table does not contain: a threshold past which a draft is "caught". There is no such number, on this page or anywhere. What exists is a texture a reader feels by the second page, and texture is editable — which is the entire premise of the scan.
Stock phrases: the vocabulary that gives drafts away
Stock phrases cluster for a boring reason: they are statistically safe. A sentence like "delving into the data reveals key insights" commits to nothing, names no finding, and fits almost any topic, which is precisely why text assembled from averages reaches for it. The problem is never one phrase — it is density. One "moreover" is a connective; six of them is a chain a grader hears by the third paragraph. The scanner counts each phrase separately because the worst offender is the one to rewrite first.
The fix that works is claim-level, not word-level. Swapping "delve into" for "examine" leaves the emptiness in place. The sentence survives only when it carries something specific: a number, a name, a comparison only your course could produce. Five before-and-after pairs (all invented, illustrative):
| Flagged draft sentence | Working rewrite |
|---|---|
| "Delving into the data reveals several key insights." | "The data answer one question plainly: students who outlined before drafting cut their revision time roughly in half." |
| "It is important to note that sleep affects memory." | "Sleep is the variable to watch here, because recall in the sleep-deprived group fell on every trial." |
| "Education plays a crucial role in shaping society." | "Compulsory schooling changed one measurable thing in the region studied: the age by which most children could read." |
| "Navigating the complexities of climate policy requires a multifaceted approach." | "Climate policy stalls on one specific fight: which industries pay first, and how much." |
| "The findings are a testament to the power of collaboration." | "The audit finished in nine days because the two halves of the team did not wait on each other." |
Each rewrite does the same two things: it deletes the costume, and it moves a real fact into the costume's place. If you cannot find the fact, that is the finding — the sentence was never about anything, and no synonym will make it about something.
Uniform sentence rhythm: the metronome effect
Human writers breathe. Sentences surge and stop; a long clause unspools, then a short one lands. Machine-assembled prose does not breathe, because each sentence is generated from the same average: twenty words, subject-first, one clause. Individually the sentences are fine. Read a paragraph of them and your ear flattens — the effect is a metronome at one speed, which readers register as synthetic even when they cannot say why.
The scanner measures it with two numbers you can check by hand. Mean sentence length is what it sounds like. Standard deviation is how far sentences stray from that mean. Divide the second by the first and you get the uniformity ratio: near 0.45 or below, the draft reads as a metronome; above 0.7, the draft already has swing. The repair is not "write shorter" — it is contrast. Follow a 30-word sentence with an 8-word one. Put a six-word verdict where a hedged 25-word sentence used to be. Here is a passage before and after (illustrative):
| Uniform (ratio ≈ 0.31) | Mixed (ratio ≈ 0.72) |
|---|---|
| "Remote work changed how teams communicate. Most companies adopted new tools within months. Employees reported mixed feelings about the change. Meetings moved to video calls almost entirely. Productivity outcomes varied across the sampled teams." | "Remote work changed how teams communicate. The tools arrived within months. What took longer was trust: managers could not see anyone working, so they scheduled more meetings, which ate the hours the commute had freed. Productivity varied. The teams that did best wrote things down." |
The second version uses the same facts. What changed is the silhouette: 32 words, then 5, then a 30-word sentence that earns its length by carrying an explanation, then 2, then 8. The cheapest test remains reading the draft aloud once — your own ear catches the metronome faster than any ratio, because your ear is the instrument the grader's eye is imitating.
Hedge stacking: when caution reads as stalling
Hedging is honest when the uncertainty is real. "The sample was small, so this suggests rather than proves" is a competent sentence — it prices the claim correctly. Hedge stacking is different: three or four softeners piled onto one claim, not because the evidence is weak but because the writer (or the generator) is hedging reflexively. "It could be argued that social media perhaps somewhat reduces study time, to some extent" commits so hard to nothing that a grader stops trusting the sentences that do commit.
The scanner counts twelve common hedges and reports them per 100 words. The banding is the tool's own heuristic, not a standard: under about 0.7 per 100 words reads as measured; creeping toward 1.8 reads as hedge-leaning; past that, the draft is stalling. The fix is rationing — one hedge per genuinely uncertain claim, chosen on purpose:
| Stacked (4 hedges, one sentence) | Rationed (1 hedge, same honesty) |
|---|---|
| "It could be argued that social media perhaps somewhat reduces study time among students, generally speaking, to some extent." | "In the forty students surveyed, social media use displaced roughly fifty minutes of weekday study time — though the sample is small." |
Note what the rationed version does with the freed space: it spends it on a number and a scope, which is the honest kind of caution — precision instead of retreat. If a hedge is load-bearing (the evidence really is thin), keep it and cut something else. If it is decorative, the sentence gets stronger the moment it leaves.
The rule of three: when every idea comes in threes
Three-part lists are the most satisfying rhythm in rhetoric, which is exactly the problem: text assembled from averages reaches for that satisfaction constantly. "X, Y, and Z" appears in every paragraph — morale, productivity, and retention; clarity, consistency, and fairness — until the pattern itself becomes a fingerprint. Human writers reach for the triad occasionally, when the material genuinely has three parts. Machine-assembled prose reaches for it because it always scans as eloquent.
The scanner counts "A, B, and C" patterns per 100 words. A handful in a 1,500-word essay is normal prose. One every hundred words or more is a tell, and the fix is editorial honesty rather than cosmetics: ask of each triad whether the third item earns its place. Two-item lists are allowed. Four-item lists are allowed when the fourth item is real. The decorative third — the item added because the list wanted a third — goes:
| Decorative triad | Edited |
|---|---|
| "The policy affects students, teachers, and parents in ways that are cultural, economic, and social." | "The policy hits teachers first: larger classes, no extra prep time, the same pay." |
| "The app is fast, intuitive, and reliable, offering speed, simplicity, and trust." | "The app loads in under a second, which is why the testers kept using it after the study ended." |
Both rewrites also broke the paired triads (two in one sentence is a special kind of tell), and both replaced the list with something checkable. That is the pattern across every fix on this page: the tell is a costume, and the edit is putting a fact where the costume was.
How to add voice: your examples, your data, your stance
Every fix above removes something. This one adds, and it is the part no scanner — and no generator — can do for you. Voice in academic writing is not personality flair; it is traceable specificity. Three levers, in the order students find easiest:
1. Your own examples
Swap the generic case for one from your actual course. Not "many companies use flexible schedules" but "the warehouse case in Week 6, where the shift-scheduling change cut overtime without cutting headcount." The second sentence is unstealable — it references a reading, a lecture, a dataset only your class has. If your essay opens with something with a pulse, the hook examples page shows openers built on specific images and stakes rather than sweeping scene-setting.
2. Concrete data
Numbers from your readings beat adjectives from nowhere. "Costs fell significantly" becomes "the subsidy cut per-unit costs from 40 to 31 dirhams in the simulated market" — invented here, but that is the shape: figure, unit, source. If your draft carries no numbers at all, that is itself a tell, and the fix is going back to your sources rather than to a thesaurus.
3. Stance
Take a position a reader could disagree with, then defend it. Hedged mush exists largely because the draft never decided anything. "The hybrid model works better for warehouse teams than for creative teams, and the evidence for that split is stronger than either side admits" is a thesis — someone could argue back, which is the point. Connectives then have real work to do; when you need them, the transition words by job page lists them by the move they make, which keeps "moreover" chains from reforming.
A before-and-after at paragraph scale (illustrative):
| Generic (voiceless) | Voiced |
|---|---|
| "Many factors influence first-year retention, and universities must address them holistically to improve outcomes." | "First-years at my college dropped out mostly in the first six weeks — the timetable clashes, the meal-plan cost, and one thing no orientation covers: not knowing anyone. The data my residence survey collected point at the third one hardest." |
The voiced version is longer on facts and shorter on posture. It could only have been written by someone who ran that survey — which is precisely what makes it safe.
The ten-minute fix pass
Run the scan, then work this list top to bottom. Tap items as you finish them — the meter tracks the pass:
- Read the fix-first list and start with its top item, not with the easiest.
- Rewrite the three worst stock-phrase sentences around facts from your course.
- Break the rhythm: split one long sentence and merge two medium ones.
- Ration hedges to one per genuinely uncertain claim.
- Cut any triad whose third item is decorative.
- Rescan, then read the draft aloud once before submitting.
What this scanner cannot do
The honest limits, because this page asks you to trust tools that state theirs:
- It does not detect AI. Nothing here estimates the probability that a machine wrote your draft. The score measures texture — formulaic habits — and nothing else.
- It false-positives on legitimate writing. Hedges can be the honest verb; passive voice is standard in Methods sections; "rather than" is correctly excluded but other misfires happen. Every flagged sentence deserves a judgment call, which is yours.
- The phrase list is curated, not exhaustive. A draft can score well and still be hollow, and a draft can flag three phrases and still be visibly yours. Absence of flags is not a certificate of quality.
- It cannot judge whether the ideas are yours. Citation, permission and integrity live entirely outside this page. If your course restricts AI assistance, the fix is authorship, not editing.
- It is one lens. Reading ease is a different lens — the readability checker measures it — and length budgets are another; the word counter handles those. Both complement this scan; neither replaces it.
On detectors themselves, keep the symmetry in mind: they are unreliable in both directions. A detector's "human" verdict proves nothing about authorship, and a "flagged" verdict proves even less — well-documented cases have flagged entirely human student writing, including text written before such tools existed. Treat every score as a prompt for a human conversation, never as evidence. If you want more tools in the same spirit, the free writing tools hub collects them, and the study guides hub holds the craft guides this page assumes.
Draft done, but it reads like a template?
If the deadline is close and the draft still feels assembled, a quoted, deadline-bound writer can rework the passages around your own material while your sources and argument stay yours. Get a free price quote and see the terms first.
Quiz: five questions on the tells
Five questions on style hygiene — what the checks count, what the fixes are, and what no tool can promise. Every wrong answer explains itself, and four options keep the guessing honest.
1. What does this scanner actually do?
2. A draft averages 24 words per sentence with a uniformity ratio of 0.31. What is the most direct fix?
3. Which sentence hides hedge stacking?
4. Why does the rule-of-three pattern ("X, Y, and Z") read as a tell when it appears in every paragraph?
5. A paragraph is flagged for stock phrases. What is the best fix?
Frequently asked questions
Is this tool an AI detector?
No — and that is deliberate. It never estimates the probability that a machine wrote your draft. It flags stylistic habits — stock phrases, uniform rhythm, hedge stacking, tidy triads, repeated openers and passive padding — that make prose feel formulaic. AI detectors are unreliable in both directions, so this page measures texture you can actually edit instead of handing you a verdict you cannot.
Can a professor tell if a draft sounds AI-written?
Professors read stacks of drafts and notice texture quickly: paragraphs the same length, sentences the same rhythm, claims hedged into mush, examples that could belong to any essay on the topic. That is a style judgment, not a detection score, and style judgments can be wrong. The practical move is the same either way: cut the tells and add specifics only you could have written.
Will removing these tells make me pass AI detectors?
Nothing on this page can promise that, and any tool that does promise it should be treated with suspicion. Detectors are unreliable in both directions — fully human writing gets flagged and machine text slips through. What removing the tells does guarantee is sturdier prose: specific claims, varied rhythm and fewer filler phrases. That is worth having regardless of what any detector says.
What triggers the robotic look most often?
Usually a pair: stock phrases plus uniform sentence rhythm. A draft can dodge every phrase on the list and still read as machine-assembled if all thirty sentences run about twenty words. Hedge stacking and rule-of-three overuse add to the effect. The scanner reports each tell separately because the fix for one — reading aloud for rhythm, cutting for hedges — does nothing for the others.
How much of my draft do I need to rewrite?
Often less than you expect, but deeper than swapping words. Replace generic examples with ones from your own course, put numbers where the draft has adjectives, and take a stance a reader could disagree with. A 200-word passage that names your data and your position beats 600 words of polished generalities — and it is the one fix no synonym swap can imitate.
Did the scan find something worth fixing?