How to Recognize AI-Generated Text by Its Style

In an editorial analysis, repeated words and constructions may catch the eye. I look not at a single turn of phrase, but at the combination of vocabulary, syntax, sentence length, and ways of linking paragraphs. Such observations describe the text, but do not, by themselves, establish its origin.

Why a text may develop a recognizable style

You cannot determine exactly how a text was prepared from the finished text alone. If a piece repeats transitions, ways of qualifying claims, or concluding remarks, I note these features. One such observation is not enough to identify the author or the tools used.

A recognizable style may be linked not to a single “favorite word.” In an editorial analysis, I look for sentences of similar length, paragraphs that follow the same logic, and a tendency to begin an explanation with a general point, then add a qualification and a practical conclusion. An individual sentence may sound natural; a series of similar sentences in one piece creates a recurring pattern. This describes a specific text; it is not a conclusion about who wrote it.

Repeated speech patterns do not, by themselves, make it possible to distinguish one way of preparing a text from another. An author's habits and a publication's requirements can be taken into account when making comparisons, but only if suitable materials are available. This article contains no such data, so the text's origin cannot be established on the basis of these features.

What is known about frequent words and expressions

The article provides no source or data that would let readers verify the figure “approximately 13 000 words and expressions” or the claim that AI systems use these words more often. Therefore, this figure and conclusions about frequency cannot be considered substantiated.

I consider a frequency marker in context. The word “extreme” may appear in a text about sports, geology, or deadlines, while the expression “This is important” may appear in an educational explanation. The examples in this article do not show how often people or specific models use these expressions.

Example	What can be claimed on the basis of the materials in this article	How to interpret it
The word “extreme”	The article provides no verifiable data about how often people or models use it.	The word itself does not prove who wrote the text.
“This is important”	The article provides no verifiable data about how often this expression is used.	You can note that a phrase is repeated in a specific text, but you cannot attribute that repetition to a particular model without supporting evidence.
“Not just X”	The article provides no verifiable comparison of how often people and models use this construction.	You can describe its repetition in a specific piece without drawing a conclusion about its origin.
“Instead of relying on X”	The article provides no verifiable comparison of how often this pattern occurs.	What is informative is the fact that the phrase is repeated in the text, not its presumed connection to a model.

When analyzing a specific piece, I note repetitions, but do not call them frequent without comparing them with other texts and describing the method. Genre and topic may be part of such a comparison: wording in an instruction and a personal message serves different purposes, but these differences do not, by themselves, point to an author. The article contains no data for numerical comparisons or attributions to specific models.

Which constructions may catch the eye

Rather than listing “forbidden” words, I note recurring relationships between parts of the text. For example, the construction “not just X” first rejects a simple interpretation and then offers a broader one. This device may be appropriate in one place; if it recurs in headings, introductions, and conclusions, the piece sounds monotonous.

The phrase “instead of relying on X” sets up a contrast between a familiar and a preferred way of acting. It may fit the meaning in an instruction, but in a particular sentence, a simple “use X” may sometimes express the idea more precisely.

The phrase “This is important” helps highlight a point when an explanation follows. If the author repeats it without naming the consequences, I note it as rhetorical padding. This observation relates to a specific text, not to a particular model.

The same introductory phrase is repeated at the beginning of many paragraphs.
Paragraphs follow the same pattern: a general point, a qualification, a warning, and a recommendation.
Symmetrical contrasts appear where the subject does not call for a contrast.
Sentences assess importance but do not state the grounds or consequences.
The text's vocabulary differs from that of other materials by the same author, if such materials are available for comparison.

I use these points to describe the text, not as a test in their own right. In my analyses, I do not consider a text's formulaic quality or editorial revisions a sufficient explanation without evidence from the specific piece. The list records observations but does not establish authorship.

Why stylistic features may vary

The article contains no data that would allow us to verify how updates to a specific model change its style and what factors affect it. When comparing two texts, an editor can note differences in sentence length, vocabulary, or composition. Without comparable source data, these differences cannot be linked to a specific cause.

The example involving the em dash requires the same caution. There is no supporting data here for the claim that Claude Opus 5.5 used it 99% less often than the previous version, Claude Opus 5. Therefore, this figure cannot be used as an established fact. In an editorial analysis, the presence or absence of an em dash is not, by itself, grounds for drawing a conclusion about a text's origin.

When analyzing a text, you can take its genre, purpose, and editorial revisions into account, but without additional information, you cannot establish which specific cause led to an observed feature. The prompts “write like a strict academic editor” and “explain the topic in simple terms” are given here as examples of different instructions, not as evidence of a specific effect on a model's style.

How to assess a text using a combination of features

I start by reading, not by counting words. I note passages that sound formulaic: identical transitions, repeated conclusions, symmetrical formulations, and sentences with excessively neat logic. Then I check whether the observations recur in different parts of the piece. Several points of similarity help describe the text's features more precisely, but do not, by themselves, establish its origin.

Save the source material. Keep the text without subsequent edits. If versions from before and after editing are available, compare them to distinguish the original features from changes that were made.
Separate observations by level. Note vocabulary, set constructions, punctuation, sentence length, and paragraph structure separately.
Compare it with the author's body of work. If the author's earlier texts are available, compare typical words, sentence length, and familiar ways of arguing. Without such material, comparison is not possible.
Check for a genre-based explanation. When comparing an educational instruction, a press release, and an advertisement, keep in mind that conclusions about style require comparable materials.
Compare the features. Note whether the frequent words, syntactic patterns, and ways of structuring paragraphs match. A match alone does not confirm machine-generated origin.
Do not replace analysis with an accusation. Without additional evidence, state a conclusion about authorship as a supposition, not as an established fact.

One marker is not enough.

What should you do if you find only an em dash or the word “extreme” in a text? Note it, but do not draw a conclusion about authorship. One such marker is not enough in an editorial analysis. The reviewer describes recurring features and takes information about editorial interventions into account, if available.

Why one marker is not enough

A single observation allows for different explanations. In a specific text, the word may have appeared because of the author's professional habit, editing, or a requirement about wording. Without additional information, it is impossible to establish which explanation is correct. Even a high frequency, if verified for a specific corpus, does not let you determine how an individual text was created based on a single word.

You cannot tell from the finished text whether the author used an automated editor or a model to prepare a draft and then revised it. Additional information about how the text was prepared is needed to describe the contributions of the author, tool, and editor.

An editorial conclusion is useful when it answers not only the question “Does the text resemble a model's response?” but also “Which specific features show this?” This formulation separates observations from conclusions about authorship and helps identify which passages need revision.

A brief conclusion for editors and teachers

In a stylistic analysis, I note repeated words, set constructions, punctuation, and recurring ways of structuring paragraphs. Claims that these features are characteristic of Claude Opus 5.5 or GPT6 Astra, as well as the numerical comparisons given earlier, are not supported by the materials in this article.

An individual marker does not establish a text's origin. Check several features, compare the work of the same author, and take genre into account, but do not present stylistic similarity as proof that an AI system was used.

Frequently asked questions

Can a frequent word prove that a text was machine-generated?

No. A frequent word may prompt further observation, but it does not prove authorship. Additional information about how the piece was prepared is needed to draw a conclusion.

Why may stylistic features differ between texts?

When comparing texts, I take genre, topic, purpose, and editorial revisions into account. The materials in this article cannot confirm that specific models caused particular differences.

Is it worth checking a text for em dashes?

You can note the punctuation mark, but cannot use it as proof. The comparison of em dash frequency across different Claude versions is not supported by the available data. Here, punctuation is treated as a feature of the text, not as an established marker of its origin.

How should you phrase a conclusion in an editorial report?

Describe your observations: which constructions recur, how the composition differs from the author's usual style, and which passages need revision. Without additional data, you cannot claim that a specific tool wrote or prepared the text.

How to Recognize AI-Generated Text by Its Style: Frequent Words and Set Phrases