How to spot AI writing (the 2026 tells and a QC checklist)
Spot AI writing fast: the vocabulary tells (delve, tapestry), the em-dash and rule-of-three rhythm, and why AI detectors misfire, plus a fix checklist.
AI writing rarely gets a fact wrong in a way that jumps out. What gives it away is texture — the vocabulary it reaches for, the rhythm of its sentences, and the way it says a lot while committing to nothing. Once you can name the tells, you see them everywhere: in cold emails, in blog posts, in the LinkedIn comment that reads like a press release. This is a practical guide to reading for those signals, not a pitch for a detector app — because as of 2026 the detectors are the least reliable part of the process.
The stakes are real for anyone who publishes. Substack now offers readers a "scan for AI text" check, platforms are demoting mass-produced AI content, and audiences have grown allergic to copy that pattern-matches to a chatbot. Whether you are vetting a freelancer's draft, checking your own output before it ships, or just trying to trust what you read, the skill is the same: spot the tells, weigh them together, and never convict on a single one.
The steps
Scan the vocabulary for AI-favored words. Large models over-reach for a small set of "impressive" words that humans rarely use in ordinary prose: delve, underscore, leverage, foster, tapestry, landscape, realm, robust, seamless, pivotal, testament, meticulous. A 2025 Science Advances study of 15+ million PubMed abstracts found the frequency of exactly these words spiked after ChatGPT launched — enough excess vocabulary to estimate at least 13.5% of 2024 abstracts were LLM-processed, and up to ~40% in some journals. Two or three of these in a short passage is a strong first flag.
Read the punctuation and sentence rhythm. AI writing is metronomic. Watch for a run of sentences all roughly the same length (15–20 words), each in tidy subject-verb-object order, with little of the fragmenting and rambling humans do when they think out loud. Em dashes are the loudest punctuation tell — models sprinkle them far more often than most writers do. The dash is not proof on its own (plenty of humans love it), but paired with uniform cadence it is telling.
Listen for the rule of three and the false-contrast crutch. AI leans hard on triads — three adjectives, three-item lists, three parallel clauses — and on the "it's not just X, it's Y" and "whether you're a… or a…" templates. Another signature is the empty pivot: "But here's the thing," "That's only half the story," "In today's fast-paced world." When every section resolves into the same rhetorical shape, a machine is filling a mold rather than making an argument.
Test for hedging and both-sides neutrality. Ask what the piece actually claims. AI hedges compulsively — "it is important to consider," "generally speaking," "there are pros and cons" — and defaults to safe, balanced framing instead of a real position. Human experts take sides, name trade-offs specifically, and are willing to be wrong. Copy that could have been written by someone with no stake in the outcome usually was, at least in part, written by a model.
Check for specificity — names, numbers, and lived detail. The hardest thing for a model to fake is concrete, first-hand detail: a real number, a dated event, a named tool with its actual quirk, an anecdote only the author could know. AI defaults to the generic ("a leading platform," "studies show," "many experts agree"). Where a human would cite the specific, AI gestures at the category. Sparse, sourceless specificity is a reliable tell — and hallucinated citations that don't resolve are a near-certainty.
Do not trust AI detectors as proof. Detectors output a confidence score, not a verdict, and they are wrong often enough to be dangerous. They false-flag human writing — especially from non-native English speakers, whose plainer phrasing scores as "AI" — and they miss lightly edited AI entirely. OpenAI shut down its own AI-text classifier in 2023 over low accuracy. Use a detector as one weak signal among the human tells above, never as the decision.
Weigh the tells together, then read once more. No single sign convicts — a good writer might love em dashes; a real expert might say "leverage." The judgment is cumulative: AI vocabulary plus flat rhythm plus rule-of-three plus hedging plus zero specific detail is a pattern, and the pattern is the evidence. Read the piece aloud last. Human writing has an uneven, breathing cadence; machine writing marches. Your ear catches what a word-list can't.
Common gotchas
One tell is not proof. Em dashes, "leverage," or a tidy triad appear in plenty of skilled human writing — only the stacking of several signals is meaningful.
The word-lists age fast. As models update and writers learn the tells, the obvious words (delve, tapestry) get trained or edited out. Rhythm, hedging, and lack of specific detail are more durable signals than any blacklist.
Lightly edited AI beats every detector and most checklists. A human who swaps the flagged words and varies a few sentences erases the surface tells while keeping the hollow structure — read for substance, not just style.
Detectors discriminate. Multiple studies show AI-text classifiers disproportionately flag non-native English writers as "AI." Never use a score to accuse a student, employee, or freelancer.
AI is confidently wrong. Fabricated statistics, fake citations, and plausible-but-false claims are a content tell in their own right — verify the facts, not just the phrasing.
Legal note
Using an AI detector to accuse or penalize someone — a student, an employee, a contractor — is risky. These tools produce probabilistic scores with documented false-positive rates and a known bias against non-native English writers, and no major detector claims courtroom-grade accuracy. Treat a detector result as a prompt to ask questions and review process, never as evidence of wrongdoing on its own.
Where Kompozy fits
Once you can spot the tells, the real use isn't policing other people's writing — it's a QC gate on your own. Kompozy is built so its output doesn't trip that checklist in the first place. Text Posts, Blog Articles, and Email Newsletters are generated against a Persona Brief that fixes voice and point of view, and run through a banned-word filter that strips the tapestry/delve/leverage vocabulary at generation time — the difference between shipping default-chatbot prose and shipping copy that sounds like you.
The practical workflow: turn this guide's tells into your review-pipeline pass. Every Kompozy generation lands in a per-post review queue before it publishes, so you read each draft for the durable signals a filter can't catch — flat rhythm, hedging, missing specifics — and add the one number, name, or lived detail that makes it unmistakably human. Then it ships across the eight primary social platforms plus blog and email on your schedule. Because Kompozy also generates net-new formats the checklist barely applies to — Persona Shorts and avatar video, Quote Graphics, Carousels — you're not just laundering text; you're diversifying away from the pure-text output that reads as AI fastest. Starter ($99/mo, 5,500 credits) covers a solo publisher who wants the brief-plus-review discipline on a lighter cadence; Pro ($299/mo, 18,000 credits) fits a team fanning reviewed copy across every platform with autopilot; Enterprise is custom. The tells are your editing standard — Kompozy just makes the drafts start closer to passing it.
Frequently asked questions
What is the single biggest tell of AI writing?
There isn't one — that's the point. The most reliable signal is a stack of them together: AI-favored vocabulary (delve, underscore, tapestry), uniform sentence rhythm, the rule of three, compulsive hedging, and a lack of specific, first-hand detail. Any one alone is weak; the combination is strong.
Are AI detectors accurate?
Not reliably. They output a probability, not a verdict, and they both false-flag human writing (especially from non-native English speakers) and miss edited AI. OpenAI discontinued its own AI-text classifier in 2023 for low accuracy. Use them as one weak signal, never as proof.
Why does AI use words like "delve" and "underscore"?
They're overrepresented in the high-quality training text and in the human feedback used to tune the models, so the model treats them as markers of good writing. A 2025 study of 15 million PubMed abstracts found these exact words surging after ChatGPT's launch — a measurable fingerprint of LLM-assisted writing.
Can you make AI writing undetectable?
You can strip the surface tells — swap the flagged words, break the rhythm, add specific detail — and beat most detectors. But that's just editing it into good writing. The durable fix isn't evasion; it's adding a real point of view, concrete detail, and facts only you could know, which is the same thing that makes writing worth reading.