A result from an AI detector can be hard to trust. Maybe the same text got a different score than before. Maybe a tool flagged a text that a person wrote.
We built Owlmark to fix both problems.
The short answer
An AI detector measures patterns in a text and decides which group the text is closer to: human writing or AI writing.
It looks at word choice and at how the words are ordered across the whole text. Then it gives a score or a verdict.
A detector does not know who wrote the text. It learned from many human and AI examples, and it measures how close the new text is to each group.
What do AI detectors look for?
A detector looks for typical AI words and constructions. It also measures how predictable the text is and how much the sentences vary.
Typical words and constructions
AI text often uses the same words and the same constructions. Common examples:
- Words such as "delve" and "tapestry".
- Phrasing such as "It's not X, it's Y".
- Ideas grouped in threes, the rule of three.
For example: "Let's delve into the tapestry of ideas. It's not a rule, it's a habit. It is clear, simple and fast." All three patterns appear in three short sentences.
Predictability
A detector checks how easy each next word is to guess. AI tends to pick likely words, so its text is often more predictable than human text.
For example, after "The sun rises in the", the word "east" is very easy to guess. After "My grandmother's kitchen was full of", many words are possible. A text where most words are easy to guess looks closer to AI.
Sentence variety
A detector also checks the rhythm of sentences. People often mix short and long sentences without planning it, and AI text is often more even.
- Even rhythm, for example: "The test was hard. The room was cold. The time was short."
- Mixed rhythm, for example: "I stopped. Then, after an hour of staring at the same page, I finally saw what the question was asking."
The statistical signature
The order of the words and how they are spread through the text differ between people and AI. This is the statistical signature that detectors measure.
Detectors compare this pattern with the patterns in their human and AI examples.
Can people spot AI writing just by reading?
Sometimes. Typical words and phrases, such as the ones in the examples above, are easy to notice. The statistical signature cannot be seen by reading.
If every typical word is cleaned out, the signature stays. So a text can look human to a reader and still be AI-generated. A good detector can catch it.
How is an AI detector trained?
A detector learns from examples of human writing and examples of AI writing. Training has three steps:
- Builders collect many texts written by people and many texts written by AI.
- The detector learns which patterns belong to each group.
- Builders tune the detector, and each team tunes it differently.
How often a detector wrongly flags human writing depends on how it was trained. For example, a detector that saw mostly casual blog posts as human writing may find a formal report strange.
Good training covers many kinds of human writing, such as news, reports, essays, casual posts and writing by non-native speakers. A detector with few kinds of human writing in its training learns a narrow idea of "human".
Why does human writing get flagged?
Many AI detectors flag human text when its style matches patterns they link to AI. Common reasons:
- Formal, structured writing. Detectors with weak training often flag formal writers.
- English as a second language. A detector trained on few kinds of English often flags it.
- Heavy editing with a tool, for example a grammar tool, can also change a text enough to be flagged.
Liang and colleagues (2023) tested seven detectors on 91 TOEFL essays by non-native English writers. The detectors flagged over half of the essays as AI-generated (study).
We built Owlmark so that human writing is almost never flagged as AI.
Why does the same text get different scores?
Each AI detector learns from its own examples and draws its own line, so two tools give two different numbers for one text. Many AI detectors even give a different number when the same text runs again.
For example, paste the same 300-word report into three tools. The results could look like this:
- Tool A shows 12% AI.
- Tool B shows 45% AI.
- Tool C shows 80% AI.
Two things explain the gap:
- Training: each tool learned from its own set of human and AI examples. Tool A may have seen many formal reports, and Tool C mostly casual posts.
- The line: each tool decides in its own place where "human" ends and "AI" starts.
Take the 45% from Tool B. A tool with its line at 70% calls that text human, and a tool with its line at 30% calls it AI. Each number is the estimate of one tool, so numbers from two tools cannot be compared.
The same tool can also give a new answer for one text. For example, a run on Monday gives 31% and a run on Tuesday gives 47%.
With Owlmark, the same text gets the same result.
Why does AI text sometimes pass?
Heavy editing by a person can move an AI text to the human side, so a detector may call it human even when AI helped to write it.
A text where a person only swapped a few typical words for plainer ones still keeps its statistical signature. When a person rewrites most of a text, the pattern changes too, and a detector may place the text on the human side.
Do AI detectors work on short texts?
Owlmark needs at least 75 words for a check. A short text has few patterns to measure. If the text is shorter, add the next paragraph and check again.
For social media, one post is often too short. Put several posts by the same account into one text and check that.
Can a text written before AI existed be flagged?
Yes. A text written long before AI existed is human writing, but a detector only measures patterns. It cannot know the date or the author of the text.
Old formal writing, such as a report from the 1990s, has the formal style that some detectors link to AI.
Can an AI detector be 100% right?
No. No AI detector is right every time, and a tool that promises 100% promises too much. OpenAI took its own AI classifier offline in July 2023 because of its low rate of accuracy (OpenAI).
We keep retraining our model. If you think we got a text wrong, email support@owlmark.app and we will review it.
What is the difference between an AI detector and a plagiarism checker?
A plagiarism checker finds text that matches existing sources, such as web pages and papers. An AI detector measures the pattern of the writing itself.
For example, copy a paragraph from a web page. A plagiarism checker finds the page.
Now ask an AI tool to write a new paragraph. No page contains it, so a plagiarism checker finds nothing, but an AI detector looks at the pattern.
The two answer different questions. A report from school is easy to sort: a plagiarism report lists matching sources, and an AI report shows a score or a verdict.
What should I do with a result?
The next step depends on why the text was checked. Find the case that fits.
Students
A teacher says the work is AI, or a detector gave a high number. Do this:
- Do not rewrite the text only to change the score. A new text hides the drafts that show how the work grew.
- Ask which tool made the report and which score it gave.
- Open the document where the text was written. In Google Docs, go to File → Version history → See version history. Word files saved to OneDrive keep a version history too.
- Show the history on screen, or send screenshots.
- Offer to explain the work in person: the main idea and why each source was chosen.
Send a short message, for example: "I wrote this myself. Can I show you the version history and explain my main argument in a short meeting?"
Next time, write in Google Docs from the start, so the history is there when needed.
Teachers and professors
Did the student write this, or did AI? A high result needs a fair next step. Follow these steps:
- Compare the text with earlier work by the same student, such as an in-class writing sample. Look at sentence length, vocabulary and the usual mistakes.
- Ask the student to open the document where the text was written and show its version history.
- Ask the student to explain the argument and the sources in a short talk. Use open questions, such as "How did you start?" and "Which source was the hardest to use?"
- Check the school's rules on AI use, so it is clear what counts as a violation.
- Record the tool, the date, the result and the parts of the text discussed.
Clients who ordered a text
The text from a paid writer or agency is in, and the question is whether a person wrote it. Try these steps:
- Ask the writer for earlier drafts or the version history of the document.
- Ask the writer to explain two choices in the text, such as why a section starts where it does.
- Ask for a short new paragraph on the same topic, so the style can be compared.
Employers and HR
A résumé or a cover letter looks generic. Was it written with AI? Invite the candidate to a short call and ask for one project from the résumé, told in their own words.
SEO and content teams
An article is waiting to be published. Check it first, and ask the author for the outline and the sources used.
FAQ
Does a high AI score mean someone cheated?
No. A high score says the pattern of the text is close to what a detector links to AI. It does not show how the text was made or whether any rule was broken.
Which languages does Owlmark work in?
English. Korean is coming soon.
Is Owlmark free?
Yes. A free account is needed to run a check. It includes 10,000 words a month, up to 1,000 words per check.
Can my school or company use Owlmark?
Schools and companies can start with the Enterprise form.