Signs of AI Writing: The Patterns That Actually Signal AI-Generated Text

Each pattern in this guide is a tendency across many texts. None of them is a fingerprint on a single document. A 2023 study published in Patterns ran seven detector tools on 91 exam essays written by non-native English speakers and on 88 essays written by U.S. eighth-graders. The detectors were near-perfect on the eighth-graders' essays, yet flagged the non-native essays as AI-generated at an average rate of 61.3%. The researchers' explanation is about word choice, not authorship. Writers with a smaller working vocabulary tend to pick more predictable words and sentence shapes. Many detectors read that same statistical property as a sign of AI generation.

The University of Glasgow's guidance for staff reaches the same conclusion. It says plainly that "academic misconduct investigations should never rely upon such software," and that the presence of a listed indicator "does not automatically mean the student has used or misused" generative AI. A match is a reason to look more closely at drafts, process, and citations. It is not a verdict.

Beyond style, a reader can check content. The Glasgow guidance quoted above also lists indicators in content and references, among them "fabricated or non-existent references" and "generic conclusions that fail to address the specific question posed."

• Sources. Check that each reference exists and says what the text claims. A 2023 study in Scientific Reports checked the citations in 84 short literature reviews by ChatGPT-3.5 and ChatGPT-4. It found that 55% of the GPT-3.5 citations and 18% of the GPT-4 citations were fabricated.
• Specifics. Look for concrete examples, data, and page references. Claims that could fit any topic are a weak signal.
• Facts and dates. Check years, statistics, and anything presented as current. The Glasgow list includes "confident presentation of plausible-sounding but inaccurate information."
• Fit. Ask whether the text answers the assignment or research question as set, rather than a broader one.
• Consistency. Compare the text with the writer's earlier work. A clear shift is a reason to ask, not evidence. Writers improve and take feedback.

These checks show where to look closer. They do not establish who wrote a text. If you use a chatbot in your own work, how to cite ChatGPT explains why facts from its answers still need a real source.

Readers often describe generated text as evenly built. Paragraphs run roughly the same length whether the point is simple or complicated. Claims and examples arrive in sets of three ("faster, cheaper, and more reliable"). Section headings follow a few stock templates, such as "Understanding X," "The Power of Y," and "X: A Complete Guide."

Repetition can be measured, and the numbers are more specific than "AI text repeats itself." A 2020 study of how language models generate text compared human writing with text from GPT-2 Large, an early model. About 0.3% of the human-written passages ended in a loop of repeated phrases. When the model always picked the single most likely next word, its text ended in such a loop more than 70% of the time. Nucleus sampling, the method the paper proposed, picks at random from a small set of likely next words. It landed close to the human rate, at around 0.4%. So these loops depend heavily on how a model picks its words, not only on the model. That makes repetition a weaker signal than the 70% figure suggests.

At the sentence level, watch for the same transitions stacked paragraph after paragraph ("Moreover," "Furthermore," "In conclusion"), for hedges piled into a single passage ("it is worth noting that," "it is important to remember," "this may vary"), and for a narrow set of words that show up far more often than they used to.

That last habit has been counted. A 2025 Science Advances study of more than 15 million PubMed abstracts tracked word frequencies from 2010 through 2024 and found specific words whose use jumped right after large language models became widely available. In 2024, the word "delves" appeared about 28 times as often as its pre-2023 trend would predict, "underscores" about 14 times as often, and "showcasing" about 11 times as often. By the study's lower-bound estimate, at least 13.5% of 2024 abstracts were processed with an LLM. The estimate differs by field, country, and journal, and reaches 40% in some of these groups.

These habits overlap with ordinary human writing for the same reason the detectors in the 2023 Patterns study misjudged non-native writers. Smooth, highly predictable word choices read as AI-like because they are statistically ordinary, and careful editing, formal academic register, and non-native English writing all produce them too. One "delves" is not a signal. Several of these words clustered in one short passage are worth a second look.

Formatting habits often mentioned include heavy use of the em dash where a comma or period would do, title-case headings where a style guide would call for sentence case, and bold type on phrases that don't need emphasis.

The evidence on em dashes is thin. Comparisons of em dash rates in generated and human text exist, but they are mostly informal, and their results change with the model and with the human texts used as a baseline. The em dash has also long been a standard mark in journalism and literary prose. Treat it as a weak signal.

A 2023 essay study in Scientific Reports compared 90 argumentative essays by students with ChatGPT essays on the same 90 topics. The ChatGPT essays came from the 22 March 2023 versions running GPT-3.5 and GPT-4. The students were not native English speakers. Their ages are unknown, but the topics point to grades 11 to 13. In ratings by 111 teachers, the generated essays scored higher for quality. The models wrote more complex sentences and used more nominalizations. These are nouns formed from verbs, such as "assessment" from "assess." The students used more modal and epistemic constructions, such as "might" or "I think." Vocabulary variety did not point one way. The students' essays were more varied than GPT-3.5's and less varied than GPT-4's.

A 2025 study in PNAS, led by Carnegie Mellon researchers, tested GPT-4o, GPT-4o Mini, and four Llama 3 models. Each model got a passage of roughly 500 words from each of 8,290 human texts in six genres, from academic writing to TV and movie scripts. The task was to write 500 more words in the same style. Four of the six models are instruction-tuned, which means they are trained to answer questions and solve problems. These four used present participial clauses, such as "leaning on his agility," at 2 to 5 times the human rate, and nominalizations at 1.5 to 2 times the human rate. The authors describe a "noun-heavy, informationally dense" style that held even when the models were prompted to match informal speech and writing.

Both studies found more nominalizations in generated text. Connecting words are less clear-cut. In the essay study, the students and GPT-3.5 did not differ significantly in their use of words such as "however" and "moreover," and GPT-4 used fewer of them than the students. Both studies tested specific models on specific kinds of writing, and the essay study's authors note that ChatGPT had already changed since the versions they tested. Findings depend on the model, the prompt, and the genre.

ChatGPT writing style

Chatbots also differ from each other. In a study presented at ICML 2025, a classifier was trained on answers from five chat models. On held-out answers, it named the right model 97.1% of the time, against 20% by chance. It did not compare chatbots with human writers. Among the five were Claude 3.5 Sonnet, Gemini 1.5 Pro, and GPT-4o, which the paper calls ChatGPT. ChatGPT favored "such as," "certainly," and "overall," and tended to bold key points and add headers. Claude favored "here," "according to," and "based on." These are tendencies of model versions tested from late 2024 to early 2025. They are not a test for one essay.

Readers do better when they look at many small signals instead of one obvious one. A human-subjects study behind a 2019 tool that color-codes how predictable each word is gave readers that word-by-word statistical view. Their rate of correctly spotting generated text rose from 54% unaided to 72%. Even with the help, roughly three in ten judgments were still wrong.

The Certentis AI Detector follows the same reasoning. It reports a document-level probability and sentence-level signals rather than a yes-or-no verdict, because statistical pattern-matching yields likelihoods, not proof. For the mechanics behind that score, and what it does and doesn't prove, see how AI detectors actually work.

If your own writing gets flagged in an academic setting, the patterns above are where a conversation starts, not a ruling. AI and academic integrity draws a practical line between acceptable AI assistance and misuse, and explains how to disclose that assistance where your institution asks for it.