It Isn't Delve: The Real Lexical Signature of Machine-Generated Text

Ask anyone how to spot AI writing and you will get a vocabulary list. Delve. Tapestry. Intricate. Realm. The words have become a meme: a shorthand so common that writers now scrub them from their own drafts out of superstition, and detectors quietly fold them into their scoring. The list feels true. So we did the obvious thing and checked it against data: 11 million words of real text, side by side, human answers and ChatGPT answers to the very same questions. The famous words barely move the needle. What actually separates machine writing from human writing is far more boring, far more durable, and hiding in the most ordinary corners of English, and the moment you see it, the vocabulary-list theory falls apart.

The test: same questions, two authors, 11 million words

To ask whether a word distinguishes machine text from human text, you need both kinds of writing answering the same prompts: otherwise you are just measuring topic. The HC3 benchmark (Hello-SimpleAI) is built for exactly this. It pairs 58,546 human answers with 26,885 ChatGPT answers across the same 24,322 questions, drawn from finance, medicine, open-domain trivia, ELI5 explanations and Wikipedia-style computer-science prompts. That is 6.55 million human tokens against 4.50 million machine tokens: enough to hear a word's true frequency rather than an accident of which questions landed in which pile.

One caveat sits over everything that follows, and it is the most newsworthy part of the whole study, so we put it up front: HC3 captures late-2022 ChatGPT. The model has turned over many times since. As you will see, that is not a footnote: it is the finding.

How you measure a "tell" without fooling yourself

Counting raw word frequencies is a trap. Common words look distinctive simply because they are common, and rare words swing wildly on a handful of occurrences. The honest tool here is a method from political science: Monroe, Colaresi and Quinn's log-odds-ratio with an informative Dirichlet prior: nicknamed "Fightin' Words," because it was built to find the language that truly separates two camps (originally Republican versus Democrat floor speeches) while controlling for how common each word is to begin with. It rewards words that one side reaches for far more than the other, and it refuses to be impressed by sheer volume. We applied it to ChatGPT-versus-human, and excluded any word appearing fewer than 80 times combined, so nothing on the list rests on noise.

Finding 1: the famous AI words are weak, or missing entirely

Here is the first surprise. The poster children of "AI writing" hardly register in early ChatGPT:

• delve, tapestry, intricate, realm: all fall below the counting threshold. They are too rare in this corpus to score at all.
• navigate appears about 4.1× more in machine text, comprehensive about 2.5×, testament a mild 1.3×.
• robust actually runs the other way: about 0.47×, meaning the humans used it more.

The viral list, in other words, is mostly describing a later model. The "delve era" everyone learned to recognize is a GPT-4-and-after phenomenon, not something intrinsic to machine text. A detector (or a personal habit, or an "AI word blacklist") tuned to that vocabulary is tuned to one season of one model. It dates instantly. (This is exactly why our own write-up of the words that read as AI treats that list as a moving target rather than a law of nature: the specific words are downstream of whichever model is in fashion.)

Finding 2: the real signature is a hedged, list-building textbook voice

So if not vocabulary fashion, then what? The strongest, most consistent separators are words so plain you would never put them on a meme. Machine text over-uses them by wide margins:

Read those words together and a register jumps out, not a thesaurus. The machine hedges (may, important), softening claims into safe, qualified statements. It scaffolds lists (including, such, also, other), perpetually enumerating, adding one more item, gesturing at completeness. And it frames everything as helpful explanation (help, overall), the cadence of a patient tutor wrapping up. Put plainly, ChatGPT in this corpus writes like a textbook with the confidence dialed down: comprehensive, balanced, faintly anxious to be useful, and structurally addicted to the bulleted summary.

This is the durable part. Vocabulary fashions come and go, but the shape of the machine voice (hedged, enumerative, explanatory) has been remarkably stable across generations, because it falls out of how these systems are trained to be helpful and harmless rather than out of any one decade's slang. The tell was never an exotic word. It was the relentless, flattening politeness of the explanatory paragraph.

And note the direction of the effect. The popular intuition is that AI writing is fancier than human writing: richer, more ornate. The data says the opposite. The machine's distinctive words are duller and more uniform; the vocabulary is flatter, not richer. It distinguishes itself by sounding like every other helpful explainer, not by reaching for the rare word.

Why a humanizer lab studied a word list

We build language tools. HumanizeMyAI is trained on a published corpus of 2,590 real student essays, and the skill underneath it (reading human writing carefully, at scale, and noticing what makes it human) is exactly the skill this question demands. A claim like "these are the words that give AI away" is squarely our problem, because half the internet now edits by it. The appeal of a word list is that it is concrete and shareable. Testing it honestly, on independent public data, with a falsifiable method and open code, is the only way to know whether the concrete thing is actually true, or just a story that feels true. It mostly isn't. The same instinct sits behind our free AI detector: read the text, not the legend. And it is the same discipline that drove the sibling to this paper, There Is No Aesthetic Surprisal Arc, which asked whether the rhythm of surprise gives AI away and found, once again, that the elegant idea loses to the boring statistics.

The honest caveat (which is also the point)

We want to be precise about what this study does and does not show, because detection is a field full of overconfident claims and we would rather under-sell.

This is HC3: one benchmark, one model generation, captured in late 2022. The specific words on both lists above will not match what a 2026 model produces. Newer systems were explicitly tuned away from the most-flagged vocabulary; that is why "delve" rose and then, on some models, began to recede again. If we re-ran this exact analysis on this year's models, the ranked words would shift.

But that instability is not a weakness of the finding: it is the finding. Any detector, any classroom rule, any "avoid these words" checklist that is calibrated to one model's vocabulary is calibrated to a thing that changes every few months. The fragile signal is the word list. The robust signal is the distributional shape (the hedged, enumerative, helpful-explainer register), which is a property of how these models are built, not of which words happened to be popular when the data was collected. A word list tells you about last season. The register tells you about the machine.

So the practical reading is the cautious one. A document full of delve and tapestry tells you less than it seems to, and might just be a careful writer who likes those words. A document that hedges every claim, enumerates every point, and reads like a balanced, faintly nervous textbook is showing you the deeper pattern, and even that is a texture to weigh, never a verdict to pronounce. The same caution applies in reverse: formal, second-language, exam-style academic prose can carry some of this register honestly, which is one more reason no single signal should ever decide the question. We say this plainly throughout, because the alternative (selling a word list as a lie detector) is how the field earned its overconfidence.

Open data, open code

This is not a black box. The complete method, the full ranked word lists for both directions, and a one-command reproduction of every number above are openly available on GitHub, with a Zenodo DOI to follow for citation. The dataset itself (HC3) is public, so anyone can pull the same 11 million words, run the same Fightin' Words analysis, and check the result rather than take it on faith. The famous AI vocabulary turns out to be a passing fashion; the durable signature is the dull, hedged, list-building voice of a textbook trying to be helpful, and we have left everything in the open so you can confirm it for yourself, and watch it shift with the next model the way the theory predicts it should.