AI Voice Patterns in Writing: What They Are and How to Spot Them

Ask someone to describe AI writing and they'll usually point at the same things: it "delves" into topics, it loves an em dash, it can't resist a bulleted list. Those aren't vague impressions — they're measurable patterns, and researchers have been counting them. Here's what the data actually shows about how AI writes, and why those patterns are so easy to spot once you know what you're looking for.

What exactly is an "AI voice pattern"?

An AI voice pattern is a word, phrase, or structural habit that large language models use far more often than human writers do — not because the words are wrong, but because the model has a narrower, more repetitive range than a real person. Humans vary their vocabulary and sentence shape across mood, audience, and energy level. A model trained to produce the statistically likely next word tends to converge on the same safe choices over and over, across millions of unrelated documents.

That convergence is what makes AI text recognizable even when every individual sentence is grammatically fine. It's not one tell — it's a cluster of small, repeated habits.

What words does AI overuse the most?

Words like "delve," "underscores," "intricate," "tapestry," and "pivotal" have spiked in frequency since ChatGPT's release, some by over 6,000%. A peer-reviewed analysis of more than 15 million PubMed abstracts found that usage of "delves" rose from 0.21 to 14.38 occurrences per million words between 2020 and 2024 — a 6,697% increase (Kobak et al. preprint, arXiv). "Underscores" rose 903.61%, from 4.50 to 45.19 per million words, and "intricate" rose 611.24%, from 6.22 to 44.22 (arXiv preprint).

A follow-up study went further and identified 21 "focal words" that ChatGPT-3.5 systematically overuses, including delve, intricate, meticulous, tapestry, realm, navigate, landscape, and pivotal (arXiv, "Why Does ChatGPT Delve So Much?").

None of these words are evidence on their own. "Delve" is a normal English word that existed long before ChatGPT. The pattern only matters in aggregate — one instance tells you nothing, a cluster of five or six in one short piece tells you something.

How much has this actually shown up in real writing?

At least 13.5% of 2024 PubMed abstracts showed signs of LLM-assisted writing, which translates to roughly 200,000 papers a year. The Kobak et al. team measured this using excess usage of style words across more than 15 million abstracts published between 2010 and 2024, then used the known publication rate (~1.5 million PubMed-indexed papers per year) to estimate volume (Kobak et al., Science Advances).

That's a specific, sourced figure for academic abstracts — not a claim about "all content on the internet," which no credible study has actually measured.

Beyond vocabulary, what else gives AI writing away?

Formatting habits matter as much as word choice — AI text uses markdown, em dashes, and bullet lists at dramatically higher rates than human writing. Pangram's analysis of millions of human and AI documents found AI text uses markdown formatting about 12x more often than human text (90 vs. 8 instances per 10,000 words), and the same 12x multiplier shows up for stock phrases like "delve into" (Pangram).

The structural tells are just as stark:

| Pattern | Human text (per 10,000 words) | AI text (per 10,000 words) | Multiplier | |---|---|---|---| | Em dashes | 2 | 17 | ~10x | | Bullet lists | 3 | 28 | ~9x | | Markdown formatting | 8 | 90 | ~12x |

There's also a subtler grammatical habit: AI models tend to avoid plain "is/are" sentences in favor of hedged constructions like "serves as a" or "marks the" — a pattern Wikipedia's editor community has documented across GPT and Gemini outputs (Wikipedia: Signs of AI Writing).

The tell isn't any single word or dash — it's density. Human writing varies its structure sentence to sentence. AI writing repeats a small set of safe structural moves (bulleted list, em-dash aside, "serves as" phrasing) at a rate 9–12x higher than people do, and that repetition is what a trained eye — or a detector — actually picks up on.

Can you reliably tell if a piece of writing was AI-generated?

Not with confidence — detectors are worse than most people assume. In controlled tests on STEM essays, the best-performing detectors (Copyleaks, GPTZero, Originality AI) reached true positive and true negative rates of 93.9% and 98.7% — genuinely strong numbers (Advances in Physiology Education). But accuracy collapses in messier real-world conditions: a Stanford-affiliated study testing seven detectors on 91 TOEFL essays written by non-native English speakers found an average false-positive rate of 61.3%, and 97.8% of essays were flagged as AI-written by at least one tool (Editage).

That gap matters. The same stylistic habits that make writing feel "foreign" or formal — simpler sentence structures, certain transition words — overlap heavily with AI's statistical habits. A detector, or a skeptical reader, can't always tell the difference between "this person learned English as a second language" and "this was generated by a model."

Has AI actually started changing how humans write?

There's early evidence of a feedback loop: human academic writing shows measurable shifts that line up with ChatGPT's rise. One analysis documented a more than 10% decrease in usage of the words "is" and "are" in academic writing in 2023 — a shift with no clear precedent in prior years (Wikipedia: Signs of AI Writing). People reading AI output constantly seem to be absorbing its habits, even when they're writing without any tool at all.

This is the real cost of AI voice patterns: it's not just that AI text sounds like AI. It's that the more we read it, the more our own writing starts to drift toward it — the hedged phrasing, the tidy bullet points, the vocabulary creep.

How do you avoid sounding like AI when you use AI to write?

The fix isn't avoiding specific words — it's giving the model something to imitate besides its own averages. A generic prompt ("write a cover letter for a marketing role") has nothing to draw on except the statistical middle of every cover letter it's seen, which is exactly how you end up with "I am excited to leverage my skills." A prompt built from your actual sentences, your actual rhythm, and a real example of something you wrote gives it a different target to aim for.

That's the gap most AI writing tools don't close — they optimize for fluency, not for sounding like a specific person. DoppelWriter is built around the second problem specifically: it learns your writing patterns instead of defaulting to the industry-average voice that triggers every tell above.

Write in your own voice at /create

FAQ

What words does AI overuse the most?

Delve, underscore, intricate, tapestry, realm, navigate, landscape, meticulous, and pivotal are among the most documented. Researchers identified 21 "focal words" that ChatGPT-3.5 used far more than typical human writing, with some rising over 6,000% in frequency in published text since 2022.

Does using words like "delve" mean a text was written by AI?

No. These are normal words that humans use too — a single word is not proof of anything. Detection research explicitly warns against flagging writing as AI-generated based on word choice alone, because it produces false accusations.

Are AI detectors reliable enough to trust?

Not for high-stakes decisions. The best detectors can hit over 93% accuracy in controlled tests, but the same category of tools falsely flagged non-native English speakers' essays as AI-written at an average rate of 61.3% in one study.

Has AI changed the way humans write?

There's early evidence of it. Researchers found a measurable drop in simple "is/are" constructions in academic writing starting in 2023, alongside a shift toward AI-style phrasing like "serves as" — a pattern with no clear precedent in writing before ChatGPT.

Why does AI-generated text use so many bullet points and em dashes?

Large language models default to a small set of structural habits that show up far more often than in human writing — bullet lists about 9x more often and em dashes about 10x more often, according to corpus analysis of millions of documents.