Signs of AI Writing: The Words, Phrases, and Patterns (2026)

The clearest signs of AI writing are not single words but clusters: a stock vocabulary ("delve", "testament", "showcasing"), a few sentence templates ("not just X, but Y", everything in threes), paragraphs of identical length under tidy headings, and confident claims with no specifics behind them. Every one of these appears in plenty of human writing. What gives a machine away is density, five or six of them inside the same 300 words, with nothing particular to the writer in between.

This guide lists the tells by category, explains why models produce them, and ends with how to fix each type. Wikipedia's editor-maintained Signs of AI writing page is the backbone of the vocabulary and structure sections; it is the most complete public catalogue and it is updated as models change.

What words does AI use a lot?

The vocabulary tell is the best documented one. Kobak and colleagues analysed more than 15 million PubMed abstracts from 2010 to 2024 and found that "style words" such as delves, underscores, and showcasing jumped in frequency right after ChatGPT's release. By their estimate at least 13.5% of 2024 biomedical abstracts had been processed with an LLM, and in some fields and countries the figure passed 40%. The words below recur across that study, the Wikipedia catalogue, and years of reader complaints, grouped by the job they're doing.

Two things most word lists miss. First, the list moves. Wikipedia's editors now split the vocabulary by model era: delve, tapestry, and testament peaked in the GPT-4 period (2023 to mid-2024); the GPT-4o era leaned on align with, fostering, and showcasing; models in use since mid-2025 lean on emphasizing, highlighting, and enhance, while "delve" has faded because every vendor heard the jokes. Second, the copula dodge is quieter and more durable than any buzzword. A text where nothing simply is anything, where every subject "serves as" or "represents", reads as machine prose even with no buzzword in it.

Before: "The museum stands as a testament to the city's vibrant cultural landscape, showcasing a tapestry of local artists." After: "The museum shows work by about forty local artists, most of them from the east side of the city."

Phrase and syntax tells

Vocabulary is easy to scrub. Sentence templates survive a synonym pass, so experienced editors learn to spot these first.

Negative parallelism ("not just X, but Y")

"It's not just a tool, it's a partner." "This isn't about speed, it's about trust." "Not only does it save time, but it also builds confidence." Wikipedia lists three variants: not just X but also Y, not X but Y, and the reversed X rather than Y. The construction lets a model sound like it is correcting a misconception nobody held. One instance is normal. Three in a page is a signature, especially as paragraph closers.

The rule of three

"Fast, scalable, and secure." "For students, professionals, and lifelong learners." Models reach for three items whether the world offers two or seven, because three sounds complete. If most lines of a bullet list contain exactly three nouns or adjectives, that's the pattern.

Em dashes in place of every other mark

ChatGPT's habit was a spaced em dash where a comma, colon, or parenthesis would do, often two per sentence ("The results — while preliminary — point to a real effect"). It became such a meme that in November 2025 Sam Altman announced ChatGPT would finally honour a custom instruction not to use them, which TechCrunch noted changes nothing about the default output. The tell is weaker in 2026 than it was, and it was always weak on its own (more below).

Signposting and throat clearing

"It's important to note that", "It is worth mentioning", "In today's fast-paced world", "Whether you're a beginner or a seasoned pro", "In conclusion", "Ultimately". These tell the reader what kind of sentence is coming instead of saying it. Humans use them too, but usually once.

Chatbot residue

Text pasted from a chat window often keeps the conversational frame: "Great question", "Certainly", "Here's a breakdown", and the sign-off "I hope this helps" or "Let me know if you'd like me to expand on any of these". Wikipedia files these under communication intended for the user. They are the strongest tell on this page because they have no other explanation.

Hedging stacks and fake candour

"While results may vary, it could potentially be argued that in some cases..." is a hedging stack: three qualifiers where one would do, or none. The newer sibling is fake candour. Jodie Cook's February 2026 Forbes list names "Honestly?" followed by nothing honest, "Here's the kicker" followed by no kicker, and the word "quietly" attached to everything ("quiet confidence", "quietly powerful"). These are the 2026 equivalents of "delve": warm-sounding set-ups the model has learned readers rate well.

Before: "Honestly? It's not just about productivity, it's about reclaiming your time, your focus, and your peace of mind." After: "I got about forty minutes a day back, mostly because I stopped rewriting the same email three times."

Structural tells

Zoom out from the sentence and AI text has a recognisable shape.

• Paragraphs of the same length. Four sentences, four sentences, four sentences. Human paragraphs run long when the idea is big and short when it isn't. Cook calls this "sentences that march in formation"; it applies to paragraphs too.
• A heading for every paragraph. Content that reads like a slide deck: H2, one paragraph, H2, one paragraph. Sometimes a heading contains nothing but the next heading.
• The restating closer. A final sentence that repeats the heading: under "Why consistency matters" the paragraph ends "That's why consistency matters." Models also love "In summary" mini-conclusions at the end of each section.
• Perfectly balanced pro and con. Every advantage has a matching disadvantage, and the text moves on without choosing. Wikipedia calls out the "Despite its [positives], [subject] faces challenges..." template with a vague "future outlook" tacked on. Cook calls it faux balance without consequence.

Content tells

These are the tells that matter most for students, researchers, and anyone whose work has to be true.

• Vague attribution. "Experts argue", "studies show", "industry reports suggest", "observers have noted". No name, no year, no link. Wikipedia treats this as a flag because it is how a model presents one source (or none) as consensus.
• No specifics at all. A 600-word piece on "improving sleep" that never names a number, a product, a person, or a time.
• Fake precision. The reverse: suspiciously neat numbers ("increases productivity by 37%") with no source. If you cannot find where a figure came from, assume the model made it up.
• Invented or broken citations. Plausible titles by real authors that don't exist, DOIs that resolve to unrelated papers, dead URLs, and links still carrying utm_source=chatgpt.com. Wikipedia editors now check DOIs and ISBNs routinely because AI drafts so often fail them. For a student, this is the tell that ends in a disciplinary meeting, not the em dash; see can professors tell if you used ChatGPT for what instructors actually check.
• Generic examples. "For example, a marketing team might use this to improve engagement." A human example names the team, the campaign, and what happened.
• Participle tails. Sentences that end with ", highlighting the importance of X" or ", reflecting a broader shift toward Y". Wikipedia calls this superficial analysis: an opinion about significance bolted onto a fact, with no source for the opinion.

Formatting tells

Formatting gives away copy-paste more than authorship, but it is where most readers first get suspicious.

• Bold first phrase on every bullet, exactly like this list. It comes from readme files and sales pages in the training data.
• Emoji in headings and as bullet markers. Rare in 2026 output, common in older text still floating around.
• Title Case In Every Heading, including ones that are full sentences.
• Curly quotes and spaced em dashes in a plain-text field, such as a code comment, a forum post, or a form that normally gets straight quotes. Word processors also produce curly quotes, so this means "pasted from somewhere formatted", not necessarily "AI".
• Horizontal rules between sections, Markdown symbols (**, #) pasted where Markdown doesn't render, and leftover artifacts such as contentReference, oaicite, turn0search0, or [cite: 1], which Wikipedia tracks by model.

Signs of AI writing in fiction

Fiction's tells are mostly about intent and texture rather than vocabulary.

• Inanimate things acting on purpose. Novelist Kevin Hess points out that models keep giving intention-verbs to things that can't intend: dawn "unfurls", a mountain "stands imperiously". The verb sounds literary and means nothing.
• Metaphors that almost land. The comparison has the rhythm of an insight but doesn't map onto the thing (Cook's phrase for it).
• Named emotion instead of shown emotion. "She felt a wave of relief wash over her" rather than anything she did.
• Every scene ends on a one-line punch, frequently a negative parallelism: "It wasn't the silence that frightened her. It was what the silence meant."
• No loose ends, no odd details. Real stories contain the irrelevant specific (the chipped mug, the uncle's bad joke). AI prose is clean in the way a stock photo is clean.

The honest caveat, which Hess himself makes: mediocre human fiction has done all of this for a century. In fiction the density rule matters even more.

Why does AI write like this?

Three plain reasons, none of them mysterious.

1. Prediction pulls toward the average. A model writes the most statistically likely next word given everything it has seen. Rare, specific facts are by definition unlikely, so they get replaced with generic, positive phrasing. Wikipedia's catalogue describes this as regression to the mean, and it explains the "no specifics" tell directly.
2. Preference tuning rewards what raters like. Chat models are fine-tuned on human demonstrations and on human rankings of model outputs (Ouyang et al., 2022). Raters tend to prefer answers that sound confident, balanced, structured, and warm. Sharma et al. found that human raters and the reward models trained on them sometimes prefer agreeable responses over accurate ones, across five assistants. That is where "Great question", the faux balance, and the reassuring "You're not alone" come from: they score well.
3. The training mix shows through. Heavy bold and bullets come from readmes and sales pages; spaced em dashes and "vibrant" come from edited magazine English and marketing copy; "In conclusion" comes from a billion student essays. And the loop now runs backwards: Yakura et al. tracked 737,083 hours of unscripted podcast speech and found that "delve", "showcase", "meticulous", and similar words rose in ordinary conversation after ChatGPT launched. People are starting to sound like the model that learned from people.

None of this is deception. It is what optimising for "sounds good to the average reader" produces.

Do em dashes (or "delve") prove a text is AI-written?

No, and acting as if they do gets real people hurt.

Single tells have high false-positive rates. When Paul Graham suggested in April 2024 that "delve" marked a pitch as ChatGPT-written, Nigerian readers pushed back that the word is everyday English in Nigeria and other countries with British-influenced schooling. Writers who have used em dashes for decades now get accused of using AI. And as the Yakura data shows, the "AI words" are migrating into ordinary speech, so the overlap grows every year.

What works is density and co-occurrence. Russell, Karpinska, and Iyyer asked annotators to judge 300 nonfiction articles written by humans, GPT-4o, Claude, or o1. A majority vote of five heavy ChatGPT users misclassified one article out of 300, beating most commercial detectors, and kept that accuracy when the AI text had been paraphrased or "humanized". Their explanations show how: they flagged vocabulary, but only as one input alongside formality, originality, and clarity. That is the model to copy. Count tells per 300 words, then ask whether there is a single sentence only this writer could have written.

Detectors measure something related but different. Automated tools score predictability and sentence-length variation rather than the tells in this guide; how AI detectors work explains what they see and where they fail. To see what a detector would make of your own draft before someone else runs it, Rewritica's free AI checker shows the score.

How to fix each category

If you are editing your own AI-assisted draft (a newsletter, a product page, anywhere it's allowed), fix by category in this order. The full method with examples is in how to humanize AI text manually.

To fix the problem at the source, these ChatGPT prompts suppress most vocabulary and template tells before the draft exists. Structural and content tells you still edit by hand, because a model cannot add a specific it doesn't have.

Frequently asked questions

What is the single most common sign of AI writing?

There isn't one reliable single sign. The most frequently reported tells are the stock vocabulary (delve, showcasing, testament), the "not just X, but Y" template, and lists of three. What actually separates AI text from human text is how many of these stack up in one passage.

Are em dashes a sign of AI writing?

Heavy, spaced em dash use was a real ChatGPT habit, but many human writers use them too, and since November 2025 ChatGPT follows a custom instruction to stop. Treat em dashes as a weak signal that only matters alongside other tells.

Which words does ChatGPT use a lot?

The classic list is delve, tapestry, testament, underscore, pivotal, crucial, intricate, meticulous, vibrant, realm, landscape, and foster. Newer models have shifted toward showcasing, highlighting, emphasizing, enhance, and align with.

How can I tell if text is AI-written without a detector?

Read for density: count how many tells appear in a 300-word stretch, then check whether the text contains anything specific to the writer (a named source, a number, a concrete example, an opinion with a reason). Polished prose with no specifics and five or more tells is the typical AI profile.

Does using these words mean my own writing will be flagged as AI?

One or two won't. A draft that combines the vocabulary, the templates, and the uniform structure may trip both human readers and detectors. Run it through an AI checker to see what a detector sees, then edit the clusters, not the individual words.

Bottom line

Treat this list as a density meter, not a blacklist. One "delve" or one em dash tells you nothing; six tells in a paragraph with no specific fact, source, or opinion tells you a lot. The tells move with every model generation, so learn the categories rather than memorising words. If you're checking someone else's text, look for content tells before style tells, because a fabricated citation matters and a three-item list doesn't. If you're checking your own, fix by category using the table above, then read the result aloud once: the machine rhythm is easier to hear than to see.