The measurable markers of AI-generated text
Since late 2022, corpus researchers, editors and platform moderators have converged on a set of measurable markers that appear more often in text produced by large language models than in text people write themselves. Some are single words counted across millions of documents; others are habits of rhythm, structure or formatting. This page catalogues the markers that have documented evidence behind them, grouped by the level of language they operate on, with the source for every claim.
Two things are true at once. The markers are real, measurable frequency shifts, confirmed in corpora as large as 15 million PubMed abstracts (Kobak et al., Science Advances 2025). And none of them is proof: every marker on this page also occurs in human writing, and what the evidence shows is a change in how often, not a fingerprint. The closing section covers what these markers cannot tell you, which matters as much as the markers themselves.
Lexical markers: the words themselves
Vocabulary is the most studied layer, because word frequencies can be counted at scale. The best evidence comes from scientific abstracts, where millions of documents with known dates allow before-and-after comparison. Note that the favoured word list shifts with model generations: the Wikipedia editors' catalogue tracks 'tapestry', 'testament', 'vibrant' and 'boasts' as tells of the 2023 to mid-2024 cohort, with later cohorts leaning instead on 'highlighting', 'showcasing' and 'enhance' (Wikipedia: Signs of AI writing).
| Marker | What it is | In AI text | Evidence | Caveat |
|---|---|---|---|---|
| Focal-word overuse | Disproportionate use of a small set of favourite words such as 'delve', 'intricate' and 'underscore'. | Frequency well above the human baseline; human writers use these words only occasionally. | Juzek and Ward identified 21 focal words with anomalously raised frequency in scientific abstracts (Juzek & Ward, COLING 2025); abrupt post-2022 spikes in style words such as 'delves', 'showcasing' and 'underscores' appear across more than 15 million PubMed abstracts, implying at least 13.5 percent of 2024 biomedical abstracts had LLM involvement (Kobak et al., Science Advances 2025). | Every focal word is ordinary English; a single occurrence carries no signal, and the word lists go stale as models change. |
| Lower lexical diversity | The ratio of distinct words to total words, a standard stylometric measure of vocabulary breadth. | Lower diversity and a smaller effective vocabulary than matched human text, with more repetitive phrasing. | Reported as a consistent finding across the studies reviewed in the survey (Terčon & Dobrovoljc, arXiv 2025). | Constrained genres such as legal or procedural writing also score low; the measure is only meaningful within a genre. |
| Part-of-speech skew | The balance of word classes: nouns, determiners and adpositions on one side, adjectives and adverbs on the other. | More nouns, determiners and adpositions and fewer adjectives and adverbs than human text, giving a dense, nominal style. | A recurring result across the corpus studies surveyed (Terčon & Dobrovoljc, arXiv 2025). | Nominal style is also the house style of academic and bureaucratic prose; and the modifiers AI text does use cluster on favourites such as 'seamlessly' and 'meticulously' (Pangram guide). |
| Elegant variation | Systematic swapping of synonyms to avoid repeating a word. | Overused, plausibly an artefact of repetition penalties in text generation; human writers repeat key terms more comfortably. | Documented as a recurring sign by the community catalogue (Wikipedia: Signs of AI writing). | Writers schooled to avoid repetition, including many non-native English speakers, show the same habit for human reasons. |
| Low specificity and stock names | A preference for generic detail over checkable specifics when content is freely generated. | Proper nouns are avoided or default to stock choices; Pangram reports that 60 to 70 percent of character names in test articles generated by ChatGPT and Claude were 'Emily' or 'Sarah'. | Reported from the vendor's internal generation tests (Pangram guide). | Applies to freely generated content; human drafts edited or polished with AI assistance keep their own specifics. |
Syntactic and rhythm markers: how sentences move
These markers live below word choice, in sentence shape and cadence. They are harder to count than single words but harder to fake away, because they reflect how the text was produced rather than which tokens were picked.
| Marker | What it is | In AI text | Evidence | Caveat |
|---|---|---|---|---|
| Uniform sentence rhythm | Variation in sentence length and shape across a passage, sometimes called burstiness. | Consistently mid-length, clause-heavy sentences with monotonous variation; human writing mixes short punches with long builds. | Described as a core phrasing difference between AI and human prose (Pangram guide), consistent with the repetitive patterning reported across surveyed studies (Terčon & Dobrovoljc, arXiv 2025). | Heavily edited corporate and technical prose is also smoothed towards uniformity. |
| Copula avoidance | Replacing plain 'is' and 'are' with dressier verbs: 'serves as', 'stands as', 'marks', 'features'. | 'X serves as Y' where a human writes 'X is Y'; the community catalogue links the pattern to a roughly 10 percent decline in 'is' and 'are' in academic writing since 2023. | Tracked with examples in the community catalogue (Wikipedia: Signs of AI writing). | That same decline in human academic prose suggests people are absorbing the habit from AI text, which weakens it as a discriminator over time. |
| Rhetorical templates | Recurring frames such as negative parallelism ('not just X, but Y'; 'it's not about X, it's about Y') and triads of three adjectives or clauses, the rule of three. | Deployed mechanically and at high frequency, often where the contrast or triad adds nothing. | Both frames are catalogued with subtypes by Wikipedia editors (Wikipedia: Signs of AI writing), and 'not only ... but also' constructions appear on Pangram's high-frequency AI phrase list (Pangram guide). | Both are classical rhetorical devices that human writers use deliberately in persuasion and speeches; density is the tell, not presence. |
| Pristine surface grammar | The absence of the small irregularities of human drafting: typos, fragments, contractions, sentences starting with 'And' or 'But'. | Spelling errors are very rare, contractions rarely used, fragments and run-ons avoided, the register uniformly formal. | Listed across Pangram's spelling-and-grammar comparison (Pangram guide). | Professionally edited text is equally clean; this only separates machine output from unedited human drafts, not from published prose. |
Discourse and structure markers: how the piece is built
Above the sentence, AI text tends to reproduce one safe essay shape regardless of topic. These markers concern what the text asserts and how it is organised end to end.
| Marker | What it is | In AI text | Evidence | Caveat |
|---|---|---|---|---|
| Significance inflation | Generic assertions that the subject matters: 'stands as a testament', 'plays a pivotal role', 'underscores its importance', 'reflects broader trends'. | Attached repeatedly to ordinary facts, often via participial add-ons ('highlighting', 'emphasizing') that assert meaning without adding analysis. | Tracked as 'undue emphasis on significance' and 'superficial analysis', with word lists (Wikipedia: Signs of AI writing). | Press releases and promotional human copy show the same tic; the density of it is what distinguishes machine output. |
| Vague authority | Attributing opinions to unnamed collectives: 'industry reports', 'experts argue', 'observers have noted', 'some critics'. | Used to imply sourcing where none exists, at a frequency well above careful human writing. | Catalogued as a recurring sign with examples (Wikipedia: Signs of AI writing). | Weasel wording long predates language models; human journalism has its own version of it. |
| Formulaic wrap-ups | A tidy closing that restates the piece, opening with 'Overall', 'In conclusion' or 'In summary', often after a 'challenges and future outlook' beat. | Conclusions run long, repeat earlier content and follow the same arc on any topic; human conclusions tend to be shorter and more particular. | Described in Pangram's structure comparison (Pangram guide) and as outline-like conclusions in the community catalogue (Wikipedia: Signs of AI writing). | The five-paragraph essay taught in schools trains humans to do exactly this. |
| Uniform paragraphing | Paragraphs of near-identical length arranged in an outline-like sequence, sometimes with bullet lists dropped into running prose. | Highly regular blocks tuned to look comprehensive; human paragraph length varies with the argument being made. | Contrasted directly with the more varied paragraphing of human writing (Pangram guide). | Templates and editorial house styles impose the same regularity on human work. |
Punctuation, formatting and tone markers
These are the most visible markers and the most contested, because they are also the easiest for humans to share and for AI users to strip out. One subset is close to conclusive: leftover machine artefacts that no human would type.
| Marker | What it is | In AI text | Evidence | Caveat |
|---|---|---|---|---|
| Em dash density | Heavy reliance on the long dash to splice explanatory asides and dramatic pauses into sentences. | Noticeably more frequent than typical human baselines, used as a default connector. | Flagged independently by both major catalogues (Wikipedia: Signs of AI writing; Pangram guide). | Many admired human stylists lean on the long dash; on its own this is among the weakest markers on this page. |
| Mechanical emphasis formatting | Boldfaced key terms, Title Case Headings, and lists built from a bold inline header, a colon, then a description. | Emphasis is applied exhaustively to every instance of a chosen term; lists appear where prose is expected; chat-interface artefacts such as raw markdown asterisks, emoji bullets or stray citation tokens ('oaicite', 'contentReference') survive pasting. | Catalogued in detail, including model-specific artefact strings (Wikipedia: Signs of AI writing). | Technical documentation legitimately uses bold-header lists and title case; the leftover artefact strings are the only near-conclusive item in this group. |
| Uniform formal-positive tone | Register: an even, impersonal formality with consistently upbeat evaluation and softened criticism. | Extremely formal unless prompted otherwise, overly positive, reluctant to criticise; the survey characterises AI text as formal and impersonal overall. | Reported in Pangram's tone comparison (Pangram guide) and as a stylistic-level finding of the survey (Terčon & Dobrovoljc, arXiv 2025). | Corporate communications train human writers into an identical register. |
What these markers cannot tell you
None of these markers is proof. Every row above describes a frequency shift, not a fingerprint: a word or habit that machine text shows more often than human text, never one that human text lacks. A single 'delve' or a tidy conclusion means nothing; even a cluster of markers only shifts probability. The Wikipedia catalogue is explicit that reliable judgement requires multiple independent signs together, and it cites evidence that general readers distinguish machine text at roughly chance, about 50 percent, with heavy LLM users reaching about 90 percent (Wikipedia: Signs of AI writing).
Base rates matter more than intuition suggests. If most of the text you check is human-written, even a marker that is genuinely more common in machine text will flag mostly humans in absolute terms. Some groups sit naturally on the wrong side of these markers: non-native English speakers taught to avoid repetition, technical and academic writers whose house style is nominal and formal, and students trained on the five-paragraph essay. Automated detectors built on these signals inherit the problem; the Wikipedia catalogue notes that detection tools carry non-trivial, documented false-positive rates (Wikipedia: Signs of AI writing).
The markers are also a moving target. The favoured vocabulary differs by model generation, so any fixed word list decays (Wikipedia: Signs of AI writing). Juzek and Ward found no evidence that focal-word overuse comes from model architecture, algorithms or training data, and point instead at human preference data used in fine-tuning, which changes with every release (Juzek & Ward, COLING 2025). And the boundary itself is drifting: human writing is absorbing machine habits, from the roughly 10 percent decline in 'is' and 'are' in academic prose since 2023 (Wikipedia: Signs of AI writing) to vocabulary change in scientific abstracts that Kobak and colleagues describe as exceeding the impact of the Covid pandemic on scientific vocabulary (Kobak et al., Science Advances 2025). A marker that works today will work less well next year, in both directions.
Methodology
This page was compiled by ScriptGrain on 2026-08-03 by reading each of the sources listed below in full on that date. A marker is included only when at least one named source documents it, and every number on the page is taken directly from the named source beside it; none originates with ScriptGrain.
Evidence quality varies and is stated per row. Peer-reviewed corpus studies (Juzek & Ward, COLING 2025; Kobak et al., Science Advances 2025) and the academic survey (Terčon & Dobrovoljc, arXiv 2025) are the strongest sources; the Wikipedia editors' catalogue is a large, actively maintained observational record; the Pangram guide is a commercial detector vendor's write-up and is labelled as such wherever it is the source.
ScriptGrain measures writing style; it sells neither AI detection nor evasion, and this page makes no detection claims. Last checked: 2026-08-03.
Sources
- Wikipedia: Signs of AI writing (community-maintained catalogue)
- Pangram, 'A Comprehensive Guide to Spotting AI Writing Patterns' (detector vendor guide)
- Terčon & Dobrovoljc, 'Linguistic Characteristics of AI-Generated Text: A Survey' (arXiv, 2025)
- Juzek & Ward, 'Why Does ChatGPT Delve So Much? Exploring the Sources of Lexical Overrepresentation in Large Language Models' (COLING 2025)
- Kobak, González-Márquez, Horvát & Lause, 'Delving into LLM-assisted writing in biomedical publications through excess vocabulary' (Science Advances, 2025; arXiv preprint)
- Florida State University News, 'Why Does ChatGPT Delve So Much? FSU researchers begin to uncover why ChatGPT overuses certain words' (press release, 2025)
ScriptGrain · Why ScriptGrain · Sample voice profiles · The Journal · Support