Skip to content

AI Writing Detectors Explained: Why They Can’t Prove Authorship

The Short Answer

An AI writing detector estimates how closely a text resembles typical output from a large language model. None of them can prove who wrote a text. They can flag human writing as machine-made, they can miss AI text that has been edited, and the organizations that build and study them have documented both problems.

That does not make pattern analysis useless. Used as a description of style, it can show a writer exactly which habits make prose sound generic. Used as a verdict on authorship, it can lead to unfair decisions. This guide explains how detectors work, what the published evidence says about their reliability, and how WritePadPro’s AI Content Detector tries to be useful without pretending to know more than it can.

How AI Detectors Work

Most detectors use one or more of three approaches.

  • Trained classifiers. A model is trained on examples of human-written and AI-generated text and learns to output a probability that a new text belongs to one group. Its accuracy depends on how closely your text resembles its training data.
  • Predictability measures. Language models choose likely next words, so their output tends to be predictable to another language model. This predictability is usually called perplexity. Low perplexity is treated as a sign of AI text, but plain, simple or formulaic human writing is predictable too.
  • Style patterns. Sentence-length variation (sometimes called burstiness), stock phrases, heavy use of transitions and constant hedging are all more common in chatbot output than in casual human writing.

Every one of these measures describes the text. None of them observes the writing process, and that is the root of the problem: the same text can come from a person or a model, and the measures cannot tell which.

What the Evidence Says About Reliability

OpenAI withdrew its own classifier

In January 2023 OpenAI released a classifier for indicating AI-written text. Its announcement, still online with an update note, reports that on a “challenge set” of English texts the classifier correctly identified 26% of AI-written text as “likely AI-written,” while labeling human-written text as AI-written 9% of the time. It also warned that the classifier was very unreliable on texts below 1,000 characters, performed significantly worse in languages other than English, and could be evaded by editing. The update note says that as of July 20, 2023 the classifier was no longer available due to its low rate of accuracy (OpenAI’s announcement, checked on 2026-10-14).

Detectors misjudged non-native English writers

A 2023 study by Liang and colleagues at Stanford, “GPT detectors are biased against non-native English writers,” ran seven widely used detectors on 91 TOEFL essays written by people for whom English is a second language. The detectors misclassified more than half of them as AI-generated, with an average false positive rate of 61.22%. All seven detectors agreed on 18 of the 91 essays (19.78%), and 89 of the 91 (97.80%) were flagged by at least one detector, while essays by US eighth-graders were judged almost perfectly. The authors also showed that simple prompting could bypass the detectors (the paper on arXiv, checked on 2026-10-14).

Why these failures happen

  • Writers with a smaller working vocabulary, and writers taught to follow templates, produce predictable text, which is exactly what detectors look for.
  • Short texts give any statistic too little to work with.
  • A few rounds of editing, by a person or another tool, can remove the patterns a detector relies on.
  • Models change. A detector trained on one generation of chatbots may not recognize the next.

How WritePadPro’s Detector Is Different

The AI Content Detector does not give you a probability that a text was written by AI. It measures eight specific writing patterns, shows the evidence for each one, and combines them into a machine-like patterns score from 0 to 100. Everything runs as ordinary JavaScript in your browser: there is no AI model, no upload and no storage.

SignalWhat it measuresWeight
Overused AI phrasesWeighted matches per 1,000 words from a list such as “a testament to,” “a myriad of” and “plays a pivotal role”×2
Sentence-length variationHow much sentence lengths vary (standard deviation ÷ mean)×1.5
Sentence-opening transitionsShare of sentences that begin with words such as “Additionally,” “Ultimately” or “Overall”×1
Repetitive openingsShare of sentences starting with the most common first word×0.75
HedgingHedges per 100 words, such as “typically,” “may” and “it can be argued”×0.75
Dashes and semicolonsEm dashes plus half-weighted semicolons per 1,000 words×0.75
Paragraph uniformityHow similar paragraph lengths are (three or more paragraphs)×0.75
Vocabulary diversityShare of different words in moving 50-word windows×0.5

The score is described as few (0–24), some (25–49), many (50–74) or strong (75–100) machine-like patterns. The tool needs at least 80 words and labels reliability by length: low under 150 words, moderate under 400, and better (but still not proof) from 400 words up. The phrase, transition and hedging lists are English only, so for other languages those signals are skipped and the page says so.

Worked Example: Same Facts, Two Styles

We wrote two versions of the same short onboarding memo for this guide and ran both through the tool. Version A follows a corporate template:

Effective onboarding plays a pivotal role in the success of every new hire. Additionally, a structured first week serves as a testament to a company’s commitment to its people. Notably, clear documentation fosters confidence and may potentially reduce early mistakes.

It can be argued that a myriad of factors typically influence how quickly new employees become productive. Mentoring, regular check-ins and well-organized resources generally help managers bolster engagement. Additionally, they showcase the values of the organization from the very first day.

Ultimately, a holistic onboarding plan is paramount. Overall, organizations that invest in the first weeks of employment are typically better positioned to retain talent and foster a positive culture over time.

Version B says the same things plainly:

New hires learn fastest when their first week has a plan. Ours fits on one page. Day one is laptop setup and lunch with the team; day two is a walk through the codebase with a mentor.

When someone gets stuck early, the cause is usually a missing or outdated document. So before anyone starts, the mentor rereads the setup guide and deletes whatever no longer applies.

We check in on day three and again at the end of week two. Fifteen minutes, one question: what slowed you down? The answers go straight into next month’s plan, and the page gets a little shorter each time.

ResultVersion AVersion B
Words / sentences112 / 8106 / 8
Overall score73: many machine-like patterns25: some machine-like patterns
Overused phrases100 (9 matches)0
Sentence-opening transitions100 (5 of 8 sentences)0
Hedging100 (6 hedges)0 (1 hedge)
Sentence-length variation8075
Paragraph uniformity100100

Three lessons come out of this. First, the score follows style, and style is something any writer can change in either direction, which is why a score cannot identify the writer. Second, even the plain version did not score zero: its three paragraphs are almost the same length and its sentences vary only moderately, so two signals still fired. Third, both texts are short, and the tool marks its reliability as low for anything under 150 words. For a real check, paste 300 words or more.

If You Are a Teacher or Editor

  • Never use a score as the only evidence. The detector’s own page says so, and the research above shows why.
  • Look at the process, not just the product. Drafts, notes, outlines and version history say far more about authorship than any statistic. WritePadPro’s editor keeps version snapshots, and most word processors keep a revision history.
  • Talk to the writer. Ask them to explain an argument, a source or a word choice. People who wrote a text can usually talk about it.
  • Consider who is writing. Non-native speakers and students trained on rigid essay templates are more likely to be flagged through no fault of their own.
  • Follow your institution’s policy on AI use and on academic integrity procedures.

If You Are a Writer Worried About a False Flag

  • Keep your drafts and notes. They are the best evidence of your own work. Writing in a tool with version history makes this automatic.
  • Write specifically. Concrete details, examples and opinions that only you could provide are the opposite of generic text. Our guide to avoiding clichés has hundreds of stock phrases to replace.
  • Vary your sentences. Mix short and long ones; the sentence length guide explains how. The Readability Improver highlights long sentences, passive voice and adverbs.
  • Cut filler transitions and hedges. Keep a transition when it signals a real turn in the argument, and keep a hedge when you are genuinely unsure.
  • Use sources honestly. If you paraphrase, cite. Our guide on how to paraphrase without plagiarizing covers the method.

Using the AI Content Detector Step by Step

  1. Paste at least 80 words. 300 or more gives the signals much more to work with.
  2. Click Analyze (or press Ctrl+Enter). You get the overall score, its band, the reliability note and a card for each of the eight signals explaining what it measured and its weight.
  3. Read the evidence. Overused phrases, hedges, sentence-opening transitions and dashes are highlighted in your text, a chart shows the length of every sentence, and the phrases found are listed with their counts. Filter the highlights by type to focus on one habit at a time.
  4. Revise, or copy the report. Copy report puts a plain-text summary on your clipboard, ending with a note that the signals describe style and cannot prove authorship.

The two built-in examples (“Example: AI-style text” and “Example: conversational text”) show the extremes in one click, which is a quick way to see what each signal responds to.

Frequently Asked Questions

Can any AI detector prove that a text was written by AI?

No. Detectors measure properties of the text, such as predictability, sentence variation and stock phrases. People can write text with those properties and edited AI text can avoid them, so a score can only describe style, never establish who wrote something.

Why did my own writing get flagged as AI?

Formal, templated or simple writing shares the patterns detectors look for. Published research found that popular detectors flagged most essays by non-native English writers as AI-generated. A flag tells you about the style of the text, not about you.

What does a score of 73 mean in WritePadPro’s detector?

It means the text shows many machine-like patterns (the 50 to 74 band), such as stock phrases, sentence-opening transitions and hedges. It is not a 73% chance that AI wrote it. Open each signal card to see exactly what was measured and where in your text it occurred.

How long does a text need to be for a meaningful result?

The tool needs at least 80 words and rates reliability as low under 150 words and moderate under 400. Statistics such as sentence-length variation and paragraph uniformity are unstable on short texts, so 300 words or more is a better minimum in practice.

Is my text uploaded when I use the AI Content Detector?

No. The analysis is plain JavaScript running in your browser. There is no AI model and no server involved, and the text is not saved or logged, so you can check drafts and confidential documents.

Does the detector work for languages other than English?

Partly. Sentence-length variation, repetitive openings, paragraph uniformity, vocabulary diversity and punctuation work for languages that separate words with spaces. The phrase, transition and hedging lists are English only, so those three signals are skipped for other languages.

Try These Tools

Related Articles