Your writing
Nothing you paste is sent anywhere. The check runs in this tab with no account and no model. Quoted text, links, code and citations are left alone.
Pattern report
This lists patterns in the wording. It is not a probability that an AI wrote the text, and there is no AI score. People write every one of these patterns.
Category breakdown
| Category | Result | What was found |
|---|
Your text, with flagged patterns
Suggestions
Tick the changes you want, pick a wording where there is a choice, then press Apply Selected Changes. Nothing changes until you do, and the original stays untouched.
Improved text
Before and after
| Measure | Before | After |
|---|
A lower count means the wording changed. It does not show that the text was written by a person.
How each result was read
Limits of this report
Eight kinds of pattern, each with its own evidence
The checker reads wording in eight categories: punctuation, stock phrases, sentence rhythm, paragraph structure, rhetorical constructions, transition openers, vocabulary and specificity. Every result says what was measured, what was seen, why a reader might notice it, how to improve it and how certain the rule is.
The categories do not add up to a score. A category level is the highest level reached by its own rules, so two measures that move together are never counted twice. Short texts get individual observations only: exact stock phrases are reported at any length, but density and rhythm readings need about 150 words.
Each pattern carries one of three evidence labels, and they are different claims. A measured shift means a published study found the word or phrase became more frequent in some text after 2022, such as biomedical abstracts. A reader report means some readers name it; nothing shows it predicts who wrote a text. A house style judgement means it is plain editorial advice. None of them is evidence that the pattern predicts who wrote a text, and the checker never says so.
What it cannot see, and what it does about that
A regular expression cannot read intent. The checker does not judge whether your claims are accurate, whether the piece has a real point of view, whether its hedging is balanced or whether its argument flows. It also does not use perplexity or burstiness, the signals detectors rely on, because they need a language model and they misjudge non-native writers.
The thresholds are provisional. They were tuned on ten books published before 1910 (one of them a 1908 translation) and a small set of texts we wrote, and they still flag about 21% of those books (63 of 300 chunks). They have not been validated on modern writing, on writers using English as an additional language, or against text of confirmed AI authorship. Dash-heavy human writing, such as Victorian prose, still trips the dash rule, because a count cannot know who wrote the text.
Two features exist because of those limits. Most flags have a Useful and Not useful button, and the answer is recorded with the pattern name only, so wrong flags can be measured. We have not yet analysed any answers, so no threshold has been changed because of them. The Kind of writing setting raises the thresholds for formal and technical prose, where linking words and stock vocabulary are normal.
Cleaning these patterns does not prevent an accusation. In our study of public accusations, 27% of challenged comments carried none of the 21 patterns measured, and the overall flag fired on 64% of ordinary long articles that nobody had challenged.
How the suggestions stay safe
Most flags get an explanation and no replacement, because fixing them depends on what you meant. Dashes get an explanation and no replacement: in blind comparisons of dash edits on modern text, readers preferred the original far more often than the edit. A replacement is offered only where it rewrites the matched words and nothing around them: removing a lead-in such as "it's important to note that", dropping a linking word at the start of a sentence, or swapping "delve into" for "look at". Numbers, names, links, quotes and citations are never touched.
Every change is optional and reversible. The original text is kept, the improved text is built from your accepted choices, and Undo or Reset to Original brings back the exact original. The checker never invents an example, a number or a personal story to make the text sound human. Add those yourself, where you have them.
The finding that several widely used AI detectors consistently misclassify non-native English writing is in GPT detectors are biased against non-native English writers. The word-frequency estimate that at least 13.5% of 2024 biomedical abstracts were processed with language models, a figure for whole collections rather than single texts, is in Delving into LLM-assisted writing in biomedical publications through excess vocabulary. The patterns readers name are catalogued on Wikipedia's Signs of AI writing page, which describes its list as signs of a possible problem and not proof that a text was machine-written.
AI writing patterns, answered
What makes writing sound AI generated?
Most people who accuse a text of being AI-written name no reason at all. When they do, the commonest reasons are that the text says very little, a feeling about its tone, or that it is off-topic. Fewer than 1 in 6 accusers named a feature of the wording or formatting, and the em dash was named by about 1 in 40, in a study of 787 public accusations. This checker reads wording patterns, sentence rhythm, repeated openings and how many concrete details a text has. People write all of these patterns and a model can avoid them, so none of them shows who wrote a text.
Are em dashes a sign of AI writing?
Not on their own. Dashes became a popular tell in online discussion, but many careful writers have always used them. This checker never treats a single dash as a signal. It reports dash use only when there are at least four and they are frequent for the length of the text, and even then it is describing the punctuation, not the author.
Can AI writing be reliably detected?
No tool can do it reliably, and this checker does not try. A 2023 study found that several widely used detectors consistently misclassified writing by non-native English speakers as AI-generated. Word-frequency research can estimate how much of a whole collection of abstracts was model-assisted, but its estimates are for collections, not for a single text.
How can I make writing sound more natural?
Replace each stock phrase with the fact it stands in for. Delete linking words that repeat logic the sentence already carries. Add a real example, number or name that you know to be true. Let short thoughts take short sentences and long thoughts take long ones. Do not add typos or slang on purpose: that makes the writing worse without changing who wrote it.
Why can human written text be flagged as AI?
Detectors guess from statistical features of the wording, and plain, formal or second-language writing shares those features with model output. Carefully edited prose also uses tidy structure and linking words. The 2023 study above found the effect for non-native English writing. This checker does not guess. It shows each pattern and lets you decide whether it is a problem.
Does my text leave my browser?
No. The analysis runs in this tab and the page sends no request that contains your text. Analytics record which buttons were used and banded counts, such as how many words fall in a size range, and never the text, the phrases found or anything you typed.
Will this make my text undetectable?
No, and it does not try to. It does not claim to get past any detector. Changing the wording changes how it reads, not who wrote it. Use the suggestions to write something clearer, and treat a lower pattern count as a style improvement only.
Need stylized text instead?
UltraTextGen turns plain text into bold, italic, cursive, and 100+ other Unicode font styles. Free and instant.
Open UltraTextGen →