Checking text for signs of machine-generated style means editorially assessing the likely involvement of artificial intelligence in creating or processing the material. This is not proof of the text's origin. An individual word, an em dash, or an AI detector result cannot reliably establish who wrote it.
A working conclusion is based on a combination of indicators. These include the text itself, revision history, drafts, sources, the nature of errors, and the author's ability to explain the decisions made.
Why an individual marker does not prove the use of AI
The Oxford comma is characteristic of English punctuation, and in a Russian text it cannot serve as an independent indicator. Em dashes appear in journalism, author columns, and literary prose long before the emergence of generative models. A word such as meaningful may reflect the author's professional vocabulary rather than a language-model prompt.
The Pew study on which the data cited in the source materials are based shows the spread of certain features across the English-language internet from approximately 2023. The researchers studied nearly 500 000 English-language web pages published over five years. In their sample, em dashes appeared approximately twice as often, Oxford commas were used 63% more often, and the construction "this is not X, this is Y" appeared almost three times as often.
These figures describe a change in the linguistic environment, not the accuracy of identifying each individual piece of material. People adopt turns of phrase from chatbot responses, editors use the same templates, and authors change their style after working with an AI model. The same indicator can appear in a completely human-written text, in a machine-generated draft after extensive editorial revision, and in material that was merely checked with AI.
What does an estimate that about 10% of English-language web pages contain signs of possible AI use mean? Only that the researchers found statistically noticeable features associated with the spread of language models. It does not mean that every tenth text was entirely created by an AI model.
Which stylistic indicators to check throughout the text
It is useful for an editor to look for a recurring set of choices rather than individual words. The longer the material, the easier it is to notice how the author begins paragraphs, connects ideas, formulates conclusions, and distributes specifics.
Vocabulary and repetition of phrasing
Among the words associated with the style of language models in the study materials are deliver, testament, interplay, crucial, meaningful, and pivotable. This list applies only to English-language texts and quickly becomes outdated. The mere presence of these words proves nothing. What should be checked is their inappropriate frequency, repetition across different sections, and replacement of simple Russian or English wording with more solemn words without any semantic necessity.
- The author repeats the same conclusion at the beginning and end of several paragraphs.
- General words such as "significant," "important," or "comprehensive" are not accompanied by measurable content.
- The text contains many connecting phrases but rarely names the person taking action, the source, or the limitation.
- Complex words appear evenly, as though distributed according to a template.
Syntax and rhythm
Machine-generated style is often noticeable in an excessively even rhythm: paragraphs are similar in size, sentences repeat the same pattern, and each section ends with a neat generalization. An editorial template can produce the same result, especially if the material was prepared by an SEO team.
Check several consecutive paragraphs. Do they begin in the same way? After two short sentences, is there a long phrase with a condition, example, or limitation? Do the transitions between ideas vary? A lively authorial text does not have to be chaotic, but it more often contains an uneven pace, specific digressions, and choices that depend on context.
Semantic specificity
Template-based rewriting often preserves surface coherence but loses detail. The text begins to contain phrases about "the importance of a comprehensive approach," "the need to take specific features into account," and "improving efficiency," although it is unclear who is acting, what exactly is changing, and by what criterion the result is evaluated.
Ask the text to answer simple questions: who performed the action, what data was it based on, under what conditions does the method not work, and what happens if there is an error? If the material confidently generalizes but does not disclose limitations, that is a signal for additional checking, not a ready-made verdict that AI was used.
How to distinguish authorial style from machine editing
AI may have been used at any stage of preparing the text. The author may have asked a model to create an outline, rewrite the introduction, reduce repetition, translate the material, or correct the grammar. After that, the final text may retain the author's facts and position while acquiring more even syntax and typical connecting phrases.
The editor compares the disputed material with verified texts by the same author. The comparison should take the topic and genre into account: a research article, product listing, and personal column require different vocabularies from the outset. If unusual constructions appear only after the author moves into a new genre, that still does not prove that an AI model was involved.
It is especially useful to check the author's decisions. Ask them to explain why a particular source was chosen, where the conclusion came from, which headline options were considered, and why one fact was left out. An author who actually worked with the material can usually reconstruct the logic of the selection even when using editing tools. Such a conversation does not replace verification, but it provides a stronger signal than counting dashes.
A practical workflow for checking text
- Save the original version. Save the disputed text without edits and mark the passages that raise questions. Do not begin by rewriting it, or the signs that need to be assessed will disappear.
- Separate observations from conclusions. Record specific facts: a construction is repeated, the vocabulary changes, links disappear, paragraphs begin in the same way. The formulation "the text looks like AI" is too vague for an editorial decision.
- Check the document history. Compare drafts, versions in the editor, comments, and the sequence of revisions. A sudden insertion of a large, coherent block without intermediate versions raises questions but does not confirm the block's source.
- Conduct a fact-check. Check names, dates, numbers, quotations, and cause-and-effect relationships against available primary materials. An AI model can produce a smooth text containing factual errors, but a factual error by itself does not prove machine generation.
- Compare the author's body of work. Take several pieces by the same author written in a similar genre and compare sentence length, connecting phrases, vocabulary, and structure. The comparison should be qualitative, not merely statistical.
- Conduct a short interview. Ask about the work process, sources, disputed wording, and passages that were edited separately. Record the answers without pressure and do not substitute them for analysis of the text.
- Formulate a limited conclusion. State what has been established, what is assumed, and what data are missing. The decision to publish, evaluate, or impose sanctions should be based on the rules of the particular editorial office, educational institution, or employer.
How to use AI detectors
An AI detector provides an additional signal about the similarity of a text to samples that its developers associate with language models. It does not reconstruct the document's history or see who wrote the original paragraph. The result depends on the language, text length, genre, topic, model version, and degree of editing.
Short texts are especially difficult to evaluate automatically. Quotations, formal style, translation, spelling corrections, and repeated professional terms affect the result. Human-written text can receive a high probability score for machine generation, while generated material may pass the check without a noticeable signal after revision.
If a detector shows a suspicious result, repeat the check on several independent passages and compare it with the revision history and fact-check. Do not use the percentage as a proven probability of authorship. It is a numerical estimate produced by a particular tool, not a measurement of the actual involvement of a person or an AI model.
How to record conclusions without issuing a categorical verdict
It is useful to apply confidence levels, defining their meaning in advance. The wording should describe the available data rather than attribute intent to the author.
| Level | What can be written | What data are needed |
|---|---|---|
| Low confidence | There are isolated stylistic similarities; the origin of the text has not been established. | Observations about vocabulary and syntax without supporting materials. |
| Medium confidence | There are several independent signals, including a sharp change in style or a history of mass insertion. | Comparison of versions, the author's body of work, and fact-checking results. |
| High confidence | Documents or an admission confirm the use of a specific tool. | Document history, metadata, working materials, or a direct explanation from the author. |
This format protects the editor from false precision. It also separates the question "Was AI used?" from the question "Were the rules violated?" One policy may allow machine translation, another may prohibit generating the main text, and a third may require disclosure of tool assistance regardless of its extent.
Why a list of indicators quickly becomes outdated
Language models change their style because developers know about characteristic words and constructions. The source materials specifically note that the word delve, previously noticeable in ChatGPT texts, has become less common in newer versions. At the same time, users adopt models' turns of phrase and spread them without any further involvement of AI.
An editorial office should revise its checklist in light of new observations, but not turn it into a catalog of prohibited words. A list of markers helps identify an issue for further checking. It does not replace drafts, sources, a conversation with the author, or an assessment of factual reliability.
Frequently asked questions
Can the use of an AI model be proven by an em dash?
No. Em dashes appear in human-written texts and depend on editorial norms, genre, and personal punctuation habits. They become useful only as one element of a broader analysis.
Does a high detector score mean that the text was written by AI?
No. Detectors make mistakes and may classify human-written text as machine-generated. Their result must be checked against document versions, the author's style, sources, and content.
What should be done if the author used AI for editing?
First, determine which rule applies to your situation. Then establish the extent of the assistance: outlining, translation, error correction, rewriting, or generation of the main text. These scenarios cannot automatically be treated as equivalent.
Is it necessary to remove all signs of machine-generated style?
No. The editorial goal is not to hunt for dashes and words, but to make the text accurate, clear, verifiable, and compliant with publication rules. Artificially removing individual markers does not correct errors or establish authorship.
Compare models before you start
The service sets its plans, limits and model catalog. If they differ from this article, contact us so we can update it and record a new review date.
Browse modelsAffiliate link: your price stays the same and the project earns a commission.
