Signs of AI writing, and how to fix each one
The patterns that make a text read as generated, from “it’s not X, it’s Y” to flat rhythm, why readers notice each one, and the edit that fixes it.
People who use chatbots for their own writing have become good at spotting chatbot prose. In a 2025 study, five of them voting together misjudged 1 article in 300. What they notice is rarely a single word. It’s a set of habits: a contrast nobody asked for, a clause like “, ensuring a seamless experience” bolted onto a fact, a list of three in every paragraph, sentences that all run the same length. Readers who can’t name those habits still feel them, and they stop trusting the page.
This is the list I edit my own drafts against. I write with AI, and every essay on this site goes through it before it’s published. Each pattern below comes with why readers notice it, how an editor fixes it, and a before and after. The free AI tells checker runs the same list on a draft in your browser and marks the lines.
How to use this list
None of these patterns proves who wrote a text. People use every one of them, often on purpose, and a careful writer working with a model can avoid them all. Treat them the way an editor would, as places to look again.
Count instead of spotting. One contrast or one dash in a 3,000-word essay is rhetoric, while three in a 150-word email are a template. What readers notice is the same move repeated, or several moves stacked in one paragraph.
Look at structure before words. Word habits come and go with model versions, and the structural ones below have been slower to change. Fix from the facts, too: swapping one buzzword for its synonym changes nothing a reader feels. Ask what the sentence is trying to claim, find the fact that would make a reader believe it, and write that. If there is no fact, cut the sentence.
Each pattern has one of the checker’s four levels. Strip means it never belongs in a finished text. Rewrite means fixing every instance, apart from the exceptions given. Cap means the device is fine once in a while and a problem when it repeats, and watch means it only matters in clusters.
Sentence patterns
“It’s not X, it’s Y”
Also “not just X, but Y”, “this isn’t about X, it’s about Y” and “less X, more Y”. The sentence sets up a view nobody stated so the next clause can knock it down and sound like an insight. It has spread. Pew Research Center counted “it’s not just X, it’s Y” at 0.87 per 10,000 words of web text in January 2023 and 2.36 in January 2026, and The Washington Post found versions of “not just X, but Y” in 6% of a sample of public ChatGPT conversations from July 2025.
Use it when the reader really holds the rejected view and saying so helps them, and then once per piece. Otherwise, state the positive claim and back it with a fact.
BeforeIt’s not just a calendar. It’s a whole new way to own your time.
AfterSlate blocks two hours for focus each morning and declines meetings that land on them.
“No X. No Y. Just Z.”
A row of absences and a payoff says what the product doesn’t do and leaves the reader to guess what it does. One plain absence next to a button is fine when it answers a real worry, such as “No card required”. Otherwise, say what does happen.
BeforeNo spreadsheets. No chasing. Just paid invoices.
AfterInvoices go out when a job is marked done, and clients get a reminder on day 7 and day 14.
“…, ensuring a smooth experience”
A participle clause bolted onto a fact to tell the reader what to think of it: “, highlighting our commitment to quality”, “, fostering collaboration”, “, making it easier than ever”. Reinhart and colleagues found instruction-tuned models using present participial clauses at two to five times the human rate, GPT-4o at 5.3 times. Cut the tail. If it held a real consequence, give that its own sentence with a concrete subject.
BeforeWe added offline mode, ensuring you stay productive wherever you are.
AfterOffline mode keeps your last 30 days of notes on the phone. Edits sync when you’re back online.
“The result? …”
A question the reader didn’t ask, answered at once, to fake suspense. The colon version works the same way: “Here’s the secret: consistency.” FAQs are exempt, and so is a headline that asks the reader’s own question in their words. Elsewhere, give the answer as a statement, with the fact that makes it interesting. The checker rates colon reveals as cap, since a colon before a list or after a label is ordinary punctuation.
BeforeThe best part? You don’t need a developer.
AfterYou can install it without a developer: paste one line into your site’s header.
“Whether you’re X or Y” and “from X to Y”
Both speak to every reader at once, so they commit to none: “whether you’re a seasoned pro or just getting started”, “everything from budgeting to big dreams”. Name the reader the text is for, or give each real group its own sentence and its own reason.
BeforeFrom solo founders to Fortune 500 teams, Pathline powers growth at every scale.
AfterMost Pathline customers are agencies with 5 to 40 people.
Stock openers, closers and filler
Openers: “In today’s fast-paced world”, “In an era where”, “When it comes to”, “It’s no secret that”, “Let’s dive in”, “Here’s the thing”, “Imagine a world where”. Closers: “In conclusion”, “Ultimately,”, “At the end of the day”, “The bottom line”, “Ready to take it to the next level?”. Filler: “a testament to”, “plays a pivotal role”, “it’s worth noting that”. Each has been used so often that readers skip it. Delete it. The first sentence should carry content, and the last should be a fact or the next step.
BeforeIn today’s fast-paced world, staying organized matters more than ever.
AfterSlate finds the next hour your whole team has free.
“Serves as” and “boasts” instead of “is” and “has”
Models dress up plain verbs to sound grander: serves as, stands as, acts as, boasts. “Is” and “has” go unnoticed, which is what you want from them.
BeforeThe Pro plan boasts unlimited projects and serves as the ideal choice for growing teams.
AfterPro has unlimited projects.
Lists of three
Adjectives in threes, slogans in threes, three bullets under every heading. Three items sound complete, so models reach for three whether or not there are three things to say. Pangram, which sells an AI detector, counts 19 lists of three per 10,000 words of AI text against 5 in human writing. The cap is for decorative lists, including three benefits that turn out to be one. A list of three different things is just a list: write it as one, and don’t break it into separate sentences to dodge the pattern.
BeforeBuilt for speed, designed for simplicity, and engineered for scale.
AfterPages render in 80 ms on a mid-range phone.
Em dashes
The tell people mention first, and the hardest to judge, because good writers have always used dashes. What gives a text away is density, and the dash as the default way into a punchline. When The Economist compared four chatbots with human journalism and fiction in July 2026, only Claude still used more em dashes than the human writers.
In body copy, aim for about one dash per 300 to 500 words, and none in short copy such as headlines, subject lines and buttons. Don’t strip them mechanically, because text with nothing but commas and periods is a newer pattern of its own. Pick the mark that fits: a comma for an aside, a colon for an explanation, a period for a real turn, parentheses for a true aside, and a dash where nothing else works. If your own writing uses dashes, keep roughly your own rate.
BeforeIt’s fast — really fast.
AfterIt opens a 200-page PDF in under a second.
Fragments for drama
“Faster. Smarter. Simpler.” One-line paragraphs for effect. A fragment can carry a fact, but three in a row are a rhythm trick. Keep one per section at most, and merge the rest into a sentence with content.
BeforeLess clutter. More clarity. Total control.
AfterHide the columns you don’t use, and the table remembers your layout on every device.
Moreover, furthermore, additionally
One sentence that opens with “Moreover,” is fine. A paragraph where every sentence opens with Moreover, Furthermore, Additionally or Notably reads as a template. Delete them. If the paragraph stops making sense without them, reorder it instead of gluing it back together.
BeforeMoreover, Slate integrates with your tools. Furthermore, it syncs across devices.
AfterSlate reads your Google and Outlook calendars, and the phone app shows the same week.
Long sentences chained with “and”
As models drop the dash, other habits show more. The Economist’s 2026 comparison found that chatbot prose used fewer commas, semicolons and parentheses than human writing. Its sentences ran longer, and “and” was the word it overused most. Split those sentences by meaning, put a real aside in parentheses, and quote a person where you have one.
Rhythm
Every sentence the same length
Sentence after sentence of 14 to 22 words, each with the same subject, verb and object. Read aloud, it drones. Asking a model for varied rhythm doesn’t fix this reliably. In a NeurIPS 2026 workshop paper, Lanzo and Angelo prompted nine model families in five ways, and showing a model examples of human rhythm gave the flattest text of all five in seven of the nine.
So measure after drafting. The checker reports the spread of sentence lengths, the standard deviation divided by the mean, for each paragraph and for the whole piece. Then restructure by meaning: give the key fact a short sentence, let the explanation run long, and merge two thin sentences into one.
BeforeOur tool helps teams coordinate meetings efficiently. It integrates with popular calendar applications seamlessly. It also provides analytics to help you understand time usage.
AfterPick a time once. Slate checks all six calendars, finds the next slot everyone has free and sends the invite before you’ve closed the tab.
Every paragraph the same size
Every paragraph three or four sentences long, every feature card the same shape, pros and cons in perfect balance. Weight by what matters to the reader instead. The feature buyers ask about gets a full paragraph with proof, and the minor ones get a line each.
A punchline at the end of every paragraph
“…because great work deserves great tools.” “…and that’s the power of simplicity.” One memorable line per piece is plenty, usually the headline or the close. End ordinary paragraphs on a fact: “…and the export is a plain CSV your accountant can open.”
The same opener again and again
Three or more sentences that start with the same word read as a tic, and the tic is often the scar of another fix, such as every dash turned into a new sentence that starts with “Then”. Change how sentences begin by restructuring them. Synonyms only move the tic.
BeforeThen we map your data. Then we import it. Then we check every row.
AfterWe map your data, import it and check every row against the source file.
Words
Words that stand in for facts
These words are fine in the right sentence. They’re a problem when they replace a fact, or when several crowd one paragraph, which is when the checker flags them. Never fix one with a synonym: “harness” for “leverage” is the same sentence. Ask what the word claims, then write the claim or cut the word.
| Family | Examples | Write instead |
|---|---|---|
| Significance puffery | testament to, pivotal role, marks a turning point, underscores the importance | The consequence for the reader, or nothing |
| Hype verbs | unlock, unleash, elevate, empower, supercharge, revolutionize, streamline, leverage | What the product does, and a result you can measure |
| Empty adjectives | seamless, robust, cutting-edge, innovative, world-class, effortless, powerful, intuitive | The evidence that would make a reader use the word |
| Grand nouns | landscape, realm, journey, ecosystem, tapestry, paradigm, synergy | The actual market, customers, tools or steps |
| Emphasis adverbs | truly, genuinely, deeply, incredibly, notably | Nothing, and if the sentence goes flat, it needed evidence |
| Brochure words | nestled, vibrant, renowned, rich heritage, a diverse array of | A concrete detail |
| Stiff verbs | utilize, facilitate, commence, conduct an analysis of | use, help, start, analyze |
| Fading favorites | delve, intricate, meticulous, showcase, underscore, foster | Plain words, though their absence proves nothing |
“Delve” is the famous case. Kobak and colleagues found “delves” 28 times more often than expected in 2024 PubMed abstracts, part of the excess vocabulary that let them estimate that at least 13.5% of that year’s abstracts were processed with a language model. Its use in arXiv abstracts then fell after early 2024.
Synonym cycling
One product called “the app”, “the platform”, “the solution” and “this powerful tool” within four sentences. Pick one name and repeat it. Readers take repetition as care, and they notice a thesaurus.
Stock idioms and morals that say nothing
“Slipping through the cracks”, “move the needle”, “low-hanging fruit”, “peace of mind”: one can be the writer’s own voice, and in a generated draft they’re filler. Name what happened instead, such as “emails sat unanswered for days”. Morals work the same way. “Communication is key” and “the right tool makes all the difference” sound like conclusions and say nothing, and rewording them keeps the padding. Cut them, or write the specific lesson if there is one.
BeforeIt’s important to note that careful planning is essential for any migration.
AfterNext time we’d export the old threads before switching.
Tone
Enthusiasm in every line
Every feature amazing, exclamation marks throughout, no trade-off anywhere. Admit one real limitation, and the rest becomes easier to believe. The checker flags each exclamation mark after the first.
BeforeBuilt to grow with you! Try it today!
AfterTeams of up to 50 use it on the free plan.
Hedges, vague sources and both sides
“It’s important to note”, “results may vary”, “studies show”, “experts agree”, “it depends on your needs”. Say the thing, or make the hedge specific: “Battery lasts about 9 hours of video, or two days of light use.” Name the study and its date, or cut the claim. And take a position: “If you have fewer than five people, a spreadsheet is fine. We’re for teams that outgrew one.”
Flattery, inflated stakes and performed candor
“Great question!”, “redefine what’s possible”, “the future of X is here”, “Let’s be honest”, “We’ve all been there”. Answer the question first, scale the claim to the actual change, and replace the intimacy with the concrete scenario it stands in for.
Content
Claims any competitor could make
Try the swap test: put a competitor’s name in the sentence, and if it’s still true, the sentence says nothing. “Enhances collaboration” and “drives efficiency” fail it. “Type @ and a name, and they get that paragraph, not the whole doc” passes. Every sentence that makes a claim needs a number, a name, a place, a date or a mechanism.
Invented proof
“Boosts productivity by 47%.” “Trusted by thousands.” Testimonials from people who don’t exist, often with names like “Dr. Sarah Chen, Head of Innovation”. A model will produce plausible proof when asked, and sometimes when it isn’t. Use real numbers and real quotes with their source, or leave a visible gap such as [NEED: number of customers] and fill it before publishing.
Padding
Restating the question, explaining why a benefit is good (“saving time matters because time is valuable”), making each point twice. Say it once, where it’s most useful. Cutting a fifth to a third of a generated draft usually loses nothing.
Formatting
Bold-label bullets
A list where every item opens with a bold label, and the label says what the item says. Write each item as one specific sentence, and keep bullets for content that is a list: steps, specs, comparisons. Arguments belong in paragraphs.
BeforeSpeed: Lightning-fast performance.
AfterSearch returns results as you type, across 50,000 notes.
Emoji as bullets
Lines and headings that start with ✅, 🚀, ✨ or 💡. People use emoji mid-sentence, where the channel calls for them.
Title Case Headings
Capitalizing every word is a style choice. Sentence case is the safer default on the web; follow your house style if it says otherwise.
Markdown that won’t render
Literal asterisks and pound signs pasted into an email, a text message or an app store listing, where they show up as characters. Format for the place the text will appear.
Invisible characters
Zero-width spaces, soft hyphens and byte-order marks can’t be seen. They break search and copying, and they often arrive with pasted text. Nothing in a finished text needs them.
Chat leftovers
Notes to the person who asked
“Here’s a revised version”, “Sure!”, “I hope this helps”, “Let me know if you’d like any changes”, “As an AI…”. Close relatives come with them: placeholders left unfilled ([Company Name], XX%, TBD, INSERT_URL), merge fields without a fallback, so Hi {first_name} reaches some readers as “Hi ,”, and the citation residue chat tools leave behind (oaicite, [cite: 3], utm_source=chatgpt.com in links). Each one shows exactly where the text came from, and none belongs in a finished text.
BeforeHere’s a revised version of your welcome email. I hope this helps!
When fixing one tell creates another
Deleting tells tends to breed new ones. In one unreviewed test of a popular humanizer skill, language-model judges flagged AI vocabulary 70 times in the original texts and 0 times in the rewrites, while their flags for a uniform beat rose from 60 to 101 and for aphoristic closers from 38 to 61. After any rewrite, look for these:
- staccato fragments in place of every long sentence
- an epigram closing each paragraph
- sections, bullets and sentences all of one size
- a performed voice: forced casualness, slang, “Look,”, invented anecdotes
- numbers that are suspiciously round or suspiciously exact, with no source
- “because”, “so” and “which means” deleted until the argument falls apart
- the same fix applied everywhere until it becomes the new tic
- punctuation reduced to commas and periods
The fix is the same as before. Go back to the facts, size sentences by meaning, keep the connectors, and stop editing lines that already work.
What isn’t a sign of AI writing
Over-correcting makes writing worse and flattens real voices. On its own, none of these is evidence:
| Feature | Why it isn’t evidence |
|---|---|
| A single em dash | A long-standing human device. Density and the dash before a punchline are the signal |
| Correct grammar, curly quotes, polish | Word processors and good writers produce them |
| Transition words such as however, so and because | Ordinary connective tissue. Only stacks of sentence openers read as a template |
| One list of three, one contrast, one rhetorical question | Classical rhetoric. The tell is doing it in every paragraph |
| Bullets for steps, specs and comparisons | That content is a list |
| Formal vocabulary, British or American spelling, contractions or none | Style and register |
| The patterns of non-native English | Taught in many school systems. A person’s English is never a tell |
| The word “delve” | Its use has fallen, and its absence proves nothing |
| An AI-detector score | See below |
Detector scores need their own warning. In a 2023 study, seven popular detectors flagged 61% of essays by non-native English speakers as AI-generated, on average, while judging essays by US eighth-graders almost perfectly. Wikipedia’s editors, who keep the best-known public list of these signs, say the same of their own list: the signs are reasons to look closer, not proof.
Check a draft
The AI tells checker runs the mechanical half of this list on a draft, in your browser: the phrase patterns, the rhythm, the punctuation and the formatting. It marks the lines and links each one back here. The rest of the list (tone, content, overcorrection) needs a reader. For Claude Code and other agents, the same list is part of my copywriter skill, which drafts from facts and runs the checker on its own output before handing it over.
Sources
- Russell, Karpinska and Iyyer, ACL 2025: people who often use ChatGPT for writing are accurate at spotting AI-generated text.
- Pew Research Center, How Much of the Internet Is Written With AI?, 20 August 2026.
- The Washington Post, analysis of public ChatGPT conversations, 13 November 2025 (paywalled).
- Reinhart et al., Do LLMs write like humans?, PNAS 2025.
- Pangram, Signs of AI writing, a detector maker’s own data.
- The Economist, on spotting AI writing, 30 July 2026 (paywalled).
- Lanzo and Angelo, Burstiness Was Measured Wrong, and Prompting Cannot Aim It, NeurIPS 2026 workshop.
- Kobak et al., Delving into LLM-assisted writing in biomedical publications through excess vocabulary, Science Advances 2025.
- Geng and Trotta, ACL Findings 2025: word choice in arXiv abstracts after ChatGPT.
- Liang et al., GPT detectors are biased against non-native English writers, Patterns 2023.
- A humanizer test, blader/humanizer issue 229, August 2026, unreviewed.
- Wikipedia, Signs of AI writing, WikiProject AI Cleanup (CC BY-SA 4.0). Its taxonomy and caveats shaped this list, written here in my own words.
- Ideas adapted from these MIT-licensed skills: coreyhaines31/marketingskills, blader/humanizer, conorbronsdon/avoid-ai-writing, petergyang/no-ai-slop and manavmishra/ZeroSlop.
A thought to add?
A good question, another perspective, or something this made you think about. There’s room for it here.
Loading the conversation…