AI writing vs human writing: the twenty linguistic measures that tell them apart

What the research found

Peer-reviewed studies since 2023 agree that machine text differs from human text in measurable ways, and that grammar and rhythm show the difference better than any single word does. Alex Reinhart and colleagues at Carnegie Mellon compared human and model text across 66 grammatical features in PNAS. In their comparison, instruction-tuned models used present participial clauses at 2 to 5 times the human rate. Vocabulary leaves a trace as well. Dmitry Kobak's team counted excess style words in PubMed and estimated that at least 13.5% of 2024 abstracts went through a language model.

The twenty measures apply that research across eight families, with rhythm and list density at either end of the sections below. Each measure produces a count from a draft.

The model and prompt affect results. Genre affects them too. So read the measures together. Wikipedia's editor guide to AI writing takes the same view: it treats a pile-up of signs as evidence, and any single sign as proof of nothing.

Rhythm measures

Rhythm measures count how much sentence and paragraph lengths vary, and burstiness is the most important of them.

How burstiness shows AI text

Burstiness is the spread of sentence lengths across a text. It's the standard deviation of sentence length divided by the mean. A human writer puts a 6-word sentence next to a 35-word one without thinking about it. Models drift toward a band of similar lengths, and that sameness is what gives a paragraph its flat, metronome feel. The first version of GPTZero, the detector Edward Tian built, scored text on perplexity and burstiness and defined burstiness as the change in predictability from one part of the document to the next. Most editing tools use sentence length as a simpler proxy for the same variation.

You can check it yourself. Count the words in each sentence of a 300-word passage. If almost every count lands within 5 words of the average, the passage has low burstiness.

Short sentences and paragraph spread

The second rhythm measure is the share of short sentences, meaning those under about 8 words. Human prose has them. Generated prose rarely does. The third is paragraph-length variation, because models build paragraphs of matching size the same way they build sentences. Genre matters for both. Legal and technical writing runs more even than an opinion column, so always compare a draft with human text of the same kind.

Vocabulary measures

Vocabulary measures count how widely a text spreads its word choices and how often it uses the words models favour.

Type-token ratio

Type-token ratio divides the number of distinct words by the total number of words. AI text scores low because a model picks the most probable next word, so the same safe words keep coming back. A study of news text by Muñoz-Ortiz and colleagues found that human articles had the highest standardized type-token ratio, at STTR of 0.491, while Falcon 7B scored an MTLD (Measure of Textual Lexical Diversity) of just 57.37 against 96.51 for humans.

Raw type-token ratio falls as a text gets longer, so a 2,000-word article always looks poorer than a 200-word one. Use a standardized version that scores fixed chunks, or use MTLD. The direction isn't fixed either. A 2026 reading-time study found that LLM rewrites had MTLD 31.83% higher than the human originals. A draft can fall outside the human range in either direction.

Marker words and word length

The fifth measure counts words that models overuse. In the Kobak data, "delves" and "showcasing" gave away LLM-assisted abstracts. "Underscores" was another marker. The sixth is average word length. The same 2026 study found that mean word length rose from 4.95 to 5.25 characters when a model rewrote human text. Each sentence takes longer to read when its words are longer and rarer.

Grammar measures

Grammar measures count four structures. The first two are present participial clauses and nominalizations. The other two cover passives and the variety of sentence types. The Reinhart study gives figures for the first three. GPT-4o used present participles at 5.3 times the rate of human writers and nominalizations 2.1 times as often. The authors' example from GPT-4o shows the habit: "Bryan, leaning on his agility, dances around the ring, evading Show's heavy blows."

Passive voice reverses that pattern. GPT-4o uses the agentless passive at half the human rate. Count it for that reason, and don't simply ban it. The tenth measure counts questions and fragments as a share of all sentences, along with the share written as imperatives. Models default to declaratives. When you check a draft, look for clauses ending in "-ing" that hang off the end of a sentence and add nothing.

Opener measures

Opener measures count how sentences begin. The first measures the share of sentences that open with the subject. The second counts repetitions of the same first word within a paragraph. Generated text opens one sentence after another with "The" or "This". The paragraph's main noun also recurs as an opener, and the subject-verb-object order repeats until it drones.

Human openers vary because people think in different shapes. A writer chooses an opener such as "When the test failed," or "But" to suit the sentence's purpose. A question provides another way to start. The check takes two minutes. Write down the first two words of every sentence in a section. If more than half start the same way, rewrite the openers before you touch anything else.

Connective measures

Connective measures count transition words per 100 sentences and the share of sentences that start with a conjunction such as And or But. So is another conjunction counted in that share. Wikipedia's guide singles out overused conjunctive phrases like "moreover" and "furthermore". Models connect paragraphs with these phrases as they generate text in sequence.

Human writers reserve transitions for turns in the logic. So a high transition density is a sign of generated text, while a sentence that starts with "But" counts toward the human side. Check for both. A draft that has "Additionally" at the head of every third paragraph needs cutting there.

Argument arc

Argument arc measures whether paragraphs and sections follow the same template. AI text repeatedly puts supporting points between a claim and a sentence that restates it. Sections end on a summary line, and the whole piece ends on "In conclusion".

The two arc measures count the share of paragraphs that follow that fixed template and the share of sections that close with a recap, while human paragraphs vary in where they end. One paragraph stops on an example or a number. Another ends on a question the next section takes up. To check a draft, read only the last sentence of each paragraph. If most of them say again what the first sentence already said, the arc is generated.

Tense measures

Tense measures count how often verb tense and grammatical person change within a text. Generated drafts sit in the present simple and third person from start to finish. Human writers use past tense for what a study found and present tense for what holds now. They move into future tense for what follows.

Person shifts work the same way. I use "you" to give the reader something to do and "I" to report what I've seen; the seventeenth and eighteenth measures track that movement. Genre changes the baseline. News leans on the past tense, and how-to pages lean on the imperative. Across genres, a page falls outside the human range if it stays in one tense or one person throughout.

List density

List density counts groups of three per 1,000 words and the share of a text written as bullets. According to Wikipedia's guide, LLMs use triplets "to make superficial analyses appear more comprehensive". Models reach for three adjectives or three benefits because that rhythm is everywhere in their training text.

Bullets are the second sign. Generated drafts break reasoning into fragments that need "because" to connect them. Human writers use a list for parallel items the reader will scan and use prose for the argument. To check, search a draft for series of exactly three joined by commas and "and". Then ask whether each bulleted block is really a list or an explanation that's been chopped up.

Human ranges per measure

A human range is a band of values drawn from human writing in the same genre; a draft passes within that band. The figures below come from published comparisons. They quantify the gap between human and AI text on the most-studied measures.

At Snoika, linguists use human-written reference text to set the target range for each of the twenty measures. These bands judge a draft against human usage in the same register and account for the wide variation among human writers. A Stanford team found that seven detectors flagged 61.3% of TOEFL essays by non-native English writers as AI-generated. James Zou, the study's senior author, explained the cause to Stanford HAI: detectors "typically score based on a metric known as 'perplexity,' which correlates with the sophistication of the writing." Check that a range's reference text matches your genre and audience before you trust it.

How Snoika scores drafts

Snoika measures whether an article reads human by scoring every draft on the twenty measures above and on sixteen AI-pattern checks, all before a human reviewer sees it. Drafts outside the human range on any measure return for revision. Because the scoring comes first, reviewers spend their time on facts and argument. The rhythm and the list habits have already been fixed.

In the evaluator's records for 666 articles scored with the same formula, the average quality score rose from 58.5 for the first articles to 82.3 in autumn 2026. Pattern points per 1,000 words dropped by about 70%, from 72.8 to 21.4. Triplets per article went from 20.4 to 3.7. Articles now have 0.5 pile-ups of "-ing" clauses, down from 9.8 per article, and em dashes have disappeared. At least 9 in 10 articles now contain no negative parallelisms; signposting filler and tailing negations are equally rare.

These are internal numbers from a single scoring formula. They show direction, not a benchmark against other tools. The numbers confirm the research's prediction that revision changes the measured features and brings drafts into the human range.

What to do next

Score your last published article on two measures before anything else: the spread of its sentence lengths and the number of three-item series per 1,000 words. Those two are the quickest to count by hand, and on generated drafts they are furthest from human values. Then select 5 human-written articles in your niche as your reference set and measure them the same way.

Your range is more useful than a general figure because the gap between AI writing vs human writing changes with the genre. Snoika scores content on these measures and tracks how brands appear in Google and AI assistants. To see where your brand stands, get a free AI visibility report.