AI checker
Paste a text and see which sentences carry the habits of AI writing, with a human-like score. Free, unlimited, English, nothing stored. The patterns are explained below the tool.
Your human-like score and the sentences that pulled it down show up here.
How it works
Two scores, one list of sentences
The human-like score (the ring, 0 to 100) comes from an open-source pattern ensemble that weighs many signals at once: vocabulary, sentence rhythm, punctuation, structure, how claims are hedged, how paragraphs open and close. Above 70 the text reads like a person wrote it; between 40 and 70 some passages read like AI; below 40 most of it does.
The style score (the badge) is ours. It starts at 100 and loses points for every habit found: stock phrases per hundred words, dashes per sentence, semicolons, uniform sentence length, lists of three, bullet lists, and the heavier patterns such as a stated lesson or a tidy closer. 100 means no stock habits found.
The highlights are what matters. Each flagged sentence shows the pattern it matched when you hover it. Fix those, and both numbers follow. The four counters under the text (stock phrases, dashes, rhythm, sentences) tell you where the points went: rhythm is the variation of sentence length, where 0.2 is monotonous and 0.5 is lively.
The patterns
What AI writing looks like, and what to do about it
Language models are trained to be helpful, balanced and polished for the average reader. The result is prose that announces its structure, inflates every point, hedges every claim and reaches for the same hundred words. None of these habits is wrong on its own. Together, in one text, they are what readers notice before they can say why. These are the families of patterns the checker looks for, with the fix we apply when we rewrite.
Overused AI words
Words that language models reach for far more often than people do. One is nothing; four in a paragraph is a fingerprint.
- delve
- tapestry
- testament
- pivotal
- multifaceted
- seamless
- leverage
- robust
- holistic
- landscape
- ever-evolving
- foster
- unlock
- elevate
- harness
- navigate
- crucial
- comprehensive
- transformative
- game-changer
- cutting-edge
Fix: Use the plain word: look at, mix, proof, key, many-sided, smooth, use, strong, whole, field, changing, help, open, raise, use, find your way, important, full.
Stock openers and filler
Phrases that announce content instead of delivering it. Readers skip them, and detectors count them.
- In today's fast-paced world
- It is important to note that
- When it comes to
- At the end of the day
- A wide range of
- First and foremost
- In conclusion
- Rest assured
Fix: Delete the opener and start with the fact. If the sentence still works, the opener was filler.
Connector openers
Sentences that begin with a logical connector, one after another, so the logic is announced rather than shown.
- Moreover,
- Furthermore,
- Additionally,
- Ultimately,
- Consequently,
- Notably,
- Indeed,
Fix: Keep one connector per paragraph at most. Put the link inside the sentence or let the order of ideas carry it.
Inflated significance
Every fact is made to matter more than it does: a step is pivotal, a detail underscores the importance, a plan sets the stage.
- stands as a testament to
- plays a crucial role in
- underscores the importance of
- marks a pivotal moment
- an enduring legacy
- the future looks bright
- setting the stage for
Fix: State the fact and let the reader judge its weight. If something is important, say why in concrete terms.
Not X but Y, and tidy closers
A contrast built for effect ("It is not about speed, it is about trust"), a one-line sentence that closes a paragraph with a flourish, or a stated lesson at the end.
- It's not just X, it's Y
- a one-line closer after a long paragraph
- The lesson is clear:
- In the end, what matters is
Fix: Cut the closer. End on the last concrete point. If the contrast is real, write both halves as plain statements.
Dashes, triads and stacked qualifiers
Em dashes used as a universal connector, lists of exactly three ("fast, simple and reliable"), and hedges piled on top of each other ("could potentially").
- an em dash in the middle of every other sentence
- clear, concise and compelling
- may potentially
- could arguably
Fix: Replace a dash with a comma, a colon, brackets or a full stop. Let lists have two items or four. Keep one hedge where you mean it.
Shallow -ing riders
A clause bolted onto the end of a sentence to add significance without adding information.
- , highlighting the need for
- , underscoring its role
- , fostering a sense of
- , showcasing the
Fix: Finish the sentence at the fact. If the rider says something new, give it its own sentence with a subject and a verb.
Sales and hype language
Brochure vocabulary that appears in AI text about anything, from a village to a database.
- nestled in the heart of
- breathtaking
- world-class
- revolutionary
- best-in-class
- unlock your full potential
- take it to the next level
- effortlessly
Fix: Describe what the thing does and what it costs. Specifics persuade; adjectives do not.
Flat rhythm
Sentences of nearly the same length, one after another. People write short ones and long ones; models settle into an even 18 to 22 words.
- five sentences in a row between 17 and 21 words
- every paragraph with exactly three sentences
Fix: Cut one sentence to five words. Let another run on. Read it aloud: where you breathe is where the full stop goes.
Formatting tics
Bulleted lists with a bold label and a colon on every line, emoji in headings, bold used as decoration, and the leftovers of a chat answer.
- **Key benefit:** ...
- 馃殌 Getting started
- I hope this helps!
- As an AI language model
- Let me know if you need anything else
Fix: Write prose where prose is expected. Use a list when the items are a list. Delete anything a chatbot would say to you.
Unnamed authority and disclaimers
"Experts agree", "studies show", "it is widely recognised" with no name, and reminders that the writer cannot know recent events.
- experts agree that
- studies have shown
- it is widely believed
- as of my last update
Fix: Name the study, the person or the source, with a year. If you cannot, drop the claim.
Why
Why models write like this
A base language model, fresh from pretraining, writes like the internet: uneven, specific, sometimes blunt. The assistant you talk to is that model after instruction tuning and preference training, where human raters rewarded answers that were complete, balanced, safe and nicely formatted. Thousands of rounds of that reward produce a house style: a preamble, three balanced points, a hedge on each, a bold label per bullet, a lesson at the end, and a vocabulary of significance (pivotal, crucial, testament) that makes every topic sound important.
Detectors learned that house style. So did readers. Research in 2026 showed that text sampled from base models, before instruction tuning, is judged human by the main commercial detectors, while the same model after tuning is caught. The habits are not a property of machines; they are a property of how assistants were trained. Which is also why they can be removed.
That is what Samesay does when you click Humanize: it runs a published method that rewrites your text with a base model and its authors' released adapter, in passes, so the assistant habits fall away while the meaning check watches your figures, names and qualifiers.
Honest limits
What this checker is not
It is not a commercial AI detector and it does not predict one. Turnitin, GPTZero, Originality, Pangram and the others are trained classifiers; they change every few months, they disagree with each other, and they are wrong on individual texts in both directions. We do not run them, we do not show their scores, and we do not promise any outcome on them. Our terms repeat that in plain words.
It is not a judge of quality. A dull paragraph with no habits scores 100. A brilliant one with two dashes and a triad scores 80. Use the score as a pointer to sentences worth a second look, not as a grade.
It is not proof of authorship, in either direction. If you are a student, the only thing that protects you is the policy of your institution: many allow AI help for some tasks and ask you to declare it. Follow it. If you are a professional, the only thing that matters is whether the text says what you meant, in your voice.
Questions
About the checker
Is this an AI detector?
No. Commercial detectors are classifiers trained on piles of human and machine text; they output a probability and say nothing about why. This checker runs a published list of writing habits over your text and shows you the sentences that carry them. It is explainable and stable, and it does not predict what Turnitin, GPTZero, Originality or any other tool will say. We never promise a result on those tools, and our terms say so.
Where do the patterns come from?
From Wikipedia's editor guide "Signs of AI writing" (maintained by WikiProject AI Cleanup, CC BY-SA 4.0), the open-source blader/humanizer catalogue that organises it into numbered patterns (the 搂 numbers in the highlights), the humanizer-stack scanners based on the StoryScope study of AI fiction (Russell et al., 2026), and the open StealthHumanizer pattern ensemble for the human-like score. Every rule is public.
Why can a text I wrote myself score low?
Because the habits are habits, not proof. Careful non-native writers, people trained on business or academic templates, and anyone who learned to write from the same textbooks as the models will trip some of these rules. Several universities have documented exactly this false positive with commercial detectors. Read the highlighted sentences: if you stand by them, keep them.
Why can a machine-written text score high?
Because a model asked for plain, specific prose produces fewer of these habits, and because a short text gives the rules little to count. A high score means we found few habits, not that a person wrote it. Nobody can tell authorship from text alone with certainty.
Is my text stored?
No. The text is scored in memory on our server and the result is sent back. Nothing is written to a database, nothing is used for training, and no third-party service sees it.
Which languages does it understand?
English. The word lists and phrase rules are English; rhythm and punctuation measures run on any Latin-script text but mean less outside English.
What are the two scores?
The ring is the human-like score, 0 to 100, from the open pattern ensemble: higher means the text looks less like typical AI output across many signals at once. The badge is the style score from our own rules: 100 means no stock habits found. They usually agree; when they do not, trust the highlighted sentences over either number.
How long a text can I check?
Up to 8,000 words at a time, 30 checks a minute. Longer documents: paste them a section at a time, the result is per sentence anyway.
What should I do with a low score?
Edit the highlighted sentences by hand, or paste the text into Samesay and let the engine rewrite it in your words, then read the diff. Either way, keep your figures, your names and your hedges: the point is a text that sounds like you, not a text that passes a test.
Found habits you want gone?
Humanize the text in your own words, then read the diff and the meaning check. 1 500 words a month free.
Humanize a draft