AI Words and Phrases — The Tells That Give Away ChatGPT

---

title: AI Words and Phrases — The Tells That Give Away ChatGPT description: The words, phrases, and punctuation habits that give away AI writing — from "delve" to the em dash — and how to fix them without over-correcting. slug: ai-words-and-phrases-tells publishedAt: "2026-07-23" author: "SynthGuard Team" category: guides tags: ["text-humanizer", "chatgpt", "ai-text", "humanization"] faq:

• q: "What words give away AI writing?" a: "Common vocabulary tells include 'delve,' 'tapestry,' 'realm,' 'testament,' 'underscore,' and 'navigate,' plus connective phrases like 'it's important to note' and 'in today's fast-paced world.' None is proof on its own, but stacked together they form a recognisable signature."
• q: "Is the em dash really a sign of AI writing?" a: "It's a weak signal. AI models use em dashes heavily and uniformly, so many em dashes deployed the same way can read as AI — but plenty of human writers love the em dash too, so it means nothing in isolation."
• q: "Do AI detectors look for these specific words?" a: "Not directly. Detectors score statistical properties like perplexity and burstiness, not a blacklist of words. But the same habits that produce these tells also produce the low-perplexity, uniform text detectors flag." related: ["make-chatgpt-sound-human", "how-to-humanize-ai-text", "does-zerogpt-actually-work"]

---

Once you've read enough ChatGPT output, you start to see the seams. Certain words appear far more often than they should. Certain phrases show up in article after article. The punctuation has a particular rhythm. Individually, none of these proves anything — but together they form a signature that experienced readers recognise instantly. This is a field guide to those tells, and to fixing them without over-correcting into a new, equally obvious pattern.

The vocabulary tells#

Language models reach for certain words because they're high-probability continuations in the polished prose the models were trained on. That's why the same handful keep surfacing.

The usual suspects:

• "Delve" — the most memed AI word of all. Humans rarely delve into anything; ChatGPT delves constantly.
• "Tapestry" — usually "rich tapestry of," applied to things that are not tapestries.
• "Realm" — "in the realm of," a fancy way to say "in."
• "Testament" — "a testament to," attached to almost anything as praise.
• "Underscore" / "underscores" — meaning "emphasises," used far more by models than people.
• "Navigate" / "navigating" — "navigating the complexities of," "navigating the landscape."
• "Landscape" — the "ever-evolving landscape," the "digital landscape."
• "Leverage," "harness," "unlock," "elevate," "robust," "seamless," "vibrant" — the marketing-adjacent cluster.

The phrase tells#

Beyond single words, models lean on connective and framing phrases that pad without adding meaning:

• "It's important to note that…"
• "In today's fast-paced world…"
• "When it comes to…"
• "At the end of the day…"
• "Plays a crucial/pivotal/vital role in…"
• "Whether you're a beginner or an expert…"
• "In conclusion" / "In summary" as a mechanical sign-off.

These are the verbal equivalent of throat-clearing. They appear because they're statistically likely transitions, and cutting them almost always tightens the writing.

The structural tells#

The subtler tells aren't words at all — they're patterns, and they matter more for detection because they drive the burstiness signal.

• Uniform sentence length. The single biggest giveaway. AI writing settles into a comfortable mid-length and stays there; human writing lurches between long and short.
• The tidy rule of three. Models love three-item lists and three-part sentences — "clear, concise, and compelling." Once you notice it, you can't unsee it.
• Symmetric paragraphs. Every paragraph roughly the same size, each with a topic sentence and neat support.
• The relentlessly even register. Never too casual, never too intense, no rough edges — polished to the point of blandness.

The em dash#

The em dash (—) deserves its own mention because it became a genuine talking point. Models use it a lot: it's a flexible, high-probability way to join related clauses, so AI text is often studded with em dashes, all deployed the same way.

Here's the honest version, though: the em dash is a weak signal. It's a legitimate, elegant piece of punctuation that many careful human writers adore. A text with em dashes isn't AI; a text with many em dashes all used identically, alongside the other tells above, leans that way. Don't purge your em dashes — just vary how you punctuate.

How to fix the tells without over-correcting#

Knowing the tells, the temptation is to run a find-and-replace and call it done. Resist that, because mechanical over-correction creates a new obvious pattern — text that conspicuously avoids every AI word reads just as unnatural as text full of them.

The better approach:

1. Cut, don't just swap. Most filler phrases should be deleted, not replaced. "It's important to note that X" is almost always just "X."
2. Replace with specifics. Where a vague AI word appears, ask what you actually mean and say that. "Navigating the complexities of pricing" becomes "figuring out what to charge."
3. Fix the rhythm. The structural tells matter most. Vary your sentence lengths deliberately — this moves the burstiness signal more than any word swap.
4. Add something real. A concrete example or a genuine opinion breaks the generic pattern better than any edit to the existing text.

Doing this by hand across a long document is tedious, especially the sentence-restructuring. A humanizer automates the mechanical part — raising burstiness, swapping predictable phrasings, normalising over-uniform punctuation — while you focus on the specifics only you can add. SynthGuard's runs its Light mode in your browser with a live burstiness meter, so you can watch the rhythm change.

For the full method — prompting, editing, and humanizing together — see how to make ChatGPT sound human and our broader guide to humanizing AI text.

The bottom line#

AI writing gives itself away through overused words ("delve," "tapestry"), padding phrases ("it's important to note"), uniform structure, and heavy, even punctuation. None is proof — detectors score statistics, not a word blacklist — but together they're a signature both readers and classifiers pick up. Fix them by cutting filler, adding specifics, and varying your rhythm, and be careful not to over-correct into a new pattern that's just as obvious. The goal isn't to scrub every trace of a word; it's to write like a person with something specific to say.