# AI Writing Tells: 13 Patterns of Structure, With Fixes (2026) > AI writing tells beyond single words: flat rhythm, lists of three, “not X, but Y”, “-ing” tails and summary endings. What each looks like and how to fix it. Source: https://hidengpt.com/guides/ai-writing-tells # What are the tells of AI writing? Updated October 10, 2026 · 8 min read · by the HidenGPT team AI writing tells are habits of shape that turn up in generated text whatever the topic: sentences of nearly equal length, lists of three, “not X, but Y” contrasts, “-ing” clauses that add commentary, “serves as” where “is” would do, unnamed experts, and paragraphs that end by summing themselves up. People use each of these too. What gives text away is several at once, with no names, numbers or dates around them. Structure survives a synonym swap, so it is a better thing to edit for than a word list. It is also where newer models still differ from people: a September 2026 study of nine models found that sentence lengths vary much less in AI articles than in human ones. | Tell | What it looks like | Why it reads as generated | Fix | |---|---|---|---| | Flat rhythm | Five sentences in a row, each 18 to 24 words long | People vary sentence length; a 2026 study found AI sentence lengths vary much less | Follow a long sentence with a short one; split one, join two | | Lists of three | “fast, simple and affordable”, or three bullet points every time | Wikipedia's editors note chatbots use them to make thin analysis look complete | Use the number of items you actually have | | “Not X, but Y” | “It's not a tool, it's a partner.” “Not just X, but also Y.” | Corrects a view nobody stated; GPT-6 Astra uses this kind of corrective phrasing about 12 times as often as human writers | State the positive claim on its own | | “Y rather than X” | “The film adapts the legend rather than retelling it.” | The same move turned around, which Wikipedia's editors find common in recent output | Keep it only if readers really believe X | | The “-ing” tail | “..., highlighting the importance of community.” | Commentary bolted onto a fact; one study found instruction-tuned models use these clauses 2 to 5 times as often as people | Cut the tail, or make it a sentence with evidence | | Dodging “is” | “serves as”, “stands as”, “boasts”, “features” | Wikipedia's editors see it replacing plain “is”, “are” and “has” | Write “is”, “are” or “has” | | Puffed-up significance | “marking a pivotal moment”, “reflects broader trends” | Claims importance instead of showing it | Give the fact that makes it matter, or cut the claim | | Flagging importance | “This matters because...”, “The distinction matters.” | “Matters because” appears about 357 times as often in GPT-6 Astra articles as in human ones | Let the example show why, and delete the flag | | Unnamed sources | “Experts argue”, “observers have noted”, “many believe” | An opinion credited to an authority nobody can check | Name the source or drop the claim | | Hedge stacks | “may potentially help”, “can provide” | GPT-6 Astra uses “may” about three times as often as GPT-4.1 | One hedge, where the doubt is real | | Summary endings | A paragraph closing on “Ultimately, this shows...”; a final “In conclusion” | Repeats what the reader just read | End on the last real point | | The challenges-and-outlook close | “Despite these challenges, X remains...”, a “Future outlook” section | A fixed template for endings, noted by Wikipedia's editors | End on a concrete next step or an open question | | Formatting habits | Title Case headings, bold on every key term, bullets with bold lead-ins | Habits from slide decks and listicles, per Wikipedia's editors | Sentence-case headings, bold once at most, prose where prose works | ## Why structure outlasts word lists Alex Reinhart and colleagues (PNAS, February 2025) gave models the first 500 words of thousands of human texts, from podcast transcripts to academic papers, and asked for the next 500 words in the same style, tone and diction. The models still wrote differently. Instruction-tuned models used present participle clauses, the “-ing” tails in the table, at 2 to 5 times the human rate, and nominalizations, nouns built from verbs such as “implementation”, at 1.5 to 2 times. GPT-4o used the passive without an agent at about half the human rate. Base models, before the tuning that turns them into assistants, sat closer to people. The practical lesson: asking a model to match your style changes its words more than its sentence shapes. Graphite, a growth agency, saw the same drift from another angle in a September 2026 study of nine models. Well-known tells such as hallmark words and boilerplate became 41% to 86% less frequent from the earliest to the latest model in each family. Yet the count of distinct patterns rose 48% across the GPT versions tested, and 55% to 72% of patterns changed between one version and the next. The tells shift between versions; they have not gone away. ## A ten-minute check on your own draft Open the draft next to a blank note and run these six checks in order. The thresholds are rules of thumb from editing, not measurements. 1. Rhythm. Count the words in ten sentences in a row. If every one falls between 15 and 25, the rhythm is flat; break it with a sentence under eight words. 2. Endings. Read only the last sentence of each paragraph. If more than half restate their paragraph, delete them and see whether anything is lost. 3. Contrasts. Search for “not”, “rather than” and “instead”. Keep a contrast only when a real reader holds the view you are correcting. 4. Tails. Look for a comma followed by a word ending in “-ing” near the end of a sentence. If the clause comments on the fact instead of adding one, cut it. 5. Lists. Count the lists of exactly three on one page. More than two is a pattern, so change one to two items or four, whichever is true. 6. Sources. Underline every claim with no name, number or date. Each needs a source, a figure from your notes, or the delete key. ## Before and after Before, 68 words in four sentences of 17, 17, 19 and 15 words: “The Alder Road garden serves as a vibrant hub for the neighborhood, fostering connection, sustainability, and well-being. It's not just a place to grow vegetables; it's a space where the whole community thrives together. Experts agree that green spaces play a crucial role in mental health, highlighting the importance of projects like this. Ultimately, the garden reflects a broader shift toward local resilience and shared responsibility in cities.” That paragraph has seven structural tells: “serves as”, a list of three, a “not just X, it's Y” contrast, unnamed experts, an “-ing” tail, a summary ending and a claim of broader significance. After, 52 words in sentences of 22, 12, 8 and 10 words: “The Alder Road garden has 24 beds, and on Saturday mornings most of them are being weeded by people who met there. Rosa, 71, grows peppers next to a family from the third floor. Two years ago it was a parking lot. The waiting list for a bed now has 19 names.” The mental health claim went because it had no named source. If it matters to the piece, it comes back with one. Every number in the second version has to come from the writer; no edit can supply them. ## What is not a tell Wikipedia's guide, written for editors looking for undisclosed AI text, lists signs that don't work: perfect grammar, a mix of casual and formal registers, bland prose, formal or academic prose in general, transition words on their own, and missing citations. On mixed registers it suggests the cause may simply be a technical background, youth, playfulness or neurodivergence. Careful writers in a second language often produce regular sentences for the same reason: they pick the safe option. The guide also says the signs point at deeper problems, such as claims nobody can verify, and warns that polishing the signs away only makes those problems harder to see. That holds for your own work. If a paragraph has no source for its main claim, varying its rhythm doesn't fix it. ## Getting a second read A person reading your draft cold is still the best check for structure, because they notice the paragraph they skimmed. For the parts that can be counted, HidenGPT's AI cliché checker runs in your browser with no sign-up and marks flat sentence rhythm, very long sentences, repeated sentence openers, strings of rhetorical questions and piles of long dashes, alongside stock phrases, with a 0-100 naturalness score. It is an editing check, not an AI detector, and it can't tell whether a claim has a source, so that part of the read stays with you. ## Common mistakes - Fixing words and leaving the skeleton: a draft with every stock word removed still reads generated if each paragraph runs claim, contrast, “-ing” tail, summary. - Breaking every list of three into two: use the number of items you really have, which is sometimes three. - Banning all contrast: one “not X, but Y” aimed at a belief your reader actually holds is good writing. - Inventing anecdotes to sound human: add only details that are true, because a made-up one is worse than none. - Judging who wrote a text from its tells: Wikipedia's editors and the Graphite authors both advise against deciding authorship from individual signs. ## Questions ### What are the most common AI writing tells? Flat sentence rhythm, lists of three, “not X, but Y” contrasts, “-ing” clauses tacked onto sentences, “serves as” in place of “is”, unnamed experts and summary endings. Single words like “delve” used to top the list, but the structural habits have lasted longer. ### What are AI writing tells in fiction? Fiction has its own set on top of the structural ones. Wikipedia's guide mentions stock names and settings that keep turning up in generated stories, such as the name “Elara Voss” and “whispering woods”; also watch for even sentence rhythm and scenes that close on a tidy reflective line. ### Are AI writing tells reliable? Not on their own. Wikipedia's editors treat them as reasons to look closer, not proof, and list several popular ones that don't work at all, such as perfect grammar or formal prose. ### Do AI writing tells change with new models? Yes. In Graphite's September 2026 study, 55% to 72% of the patterns differed between consecutive versions of the same model family, and GPT-6 Astra showed slightly more distinct tells than GPT-5.6 Sol. ### Is the rule of three a sign of AI writing? Only when it repeats. Lists of three are an old rhetorical device; Wikipedia's editors note that chatbots overuse them, especially in places where most people wouldn't bother with a flourish. ### Why does AI write “it's not X, it's Y”? Nobody has shown why, but it can be counted: Graphite found GPT-6 Astra using corrective phrases such as “rather than” and “not simply” about 12 times as often as human writers. The construction implies the reader was about to get something wrong, which is why it grates. ### How do I remove AI tells from my own writing? Edit for shape before words. Vary sentence length, cut paragraph-ending summaries, drop contrasts nobody needs, turn “-ing” tails into facts or delete them, name every source, then read the draft aloud. ## Sources - [Wikipedia: Signs of AI writing (editors' field guide)](https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing) (checked October 10, 2026) - [Reinhart et al., Do LLMs write like humans? Variation in grammatical and rhetorical styles, PNAS (2025)](https://pmc.ncbi.nlm.nih.gov/articles/PMC11874169/) (checked October 10, 2026) - [Graphite research: AI Tells (16 September 2026)](https://graphite.io/five-percent/research/ai-tells) (checked October 10, 2026) ## Related - [I wrote this myself, but it sounds like AI. What am I doing that reads that way?](https://hidengpt.com/write/my-writing-sounds-like-ai) - [My draft reads like ChatGPT wrote it. How do I make it sound like a person?](https://hidengpt.com/write/make-ai-writing-sound-human) - [How do I write an essay that doesn't sound like AI?](https://hidengpt.com/write/essay-that-doesnt-sound-like-ai) - [An AI detector says my essay is AI, but I wrote it. What do I do?](https://hidengpt.com/write/ai-detector-flagged-my-essay) - [What words does ChatGPT use a lot?](https://hidengpt.com/guides/words-chatgpt-uses-a-lot) - [Why does ChatGPT use so many em dashes?](https://hidengpt.com/guides/why-does-chatgpt-use-em-dashes) - [How do I make ChatGPT sound less like AI?](https://hidengpt.com/guides/how-to-make-chatgpt-sound-less-like-ai) - [Do AI detectors work?](https://hidengpt.com/guides/do-ai-detectors-work) - [Does my writing sound like AI?](https://hidengpt.com/tools/ai-cliche-checker) [Add HidenGPT as a preferred source on Google](https://www.google.com/preferences/source?q=hidengpt.com) to see these guides more often in search and AI Overviews.