Shlokah Research · July 2026
What statistically gives AI writing away.
We ran our deterministic tell-analysis engine over 8,000 passages: 4,000 written by humans and 4,000 by ChatGPT, drawn from the public HC3 corpus. No probability guesses, no black-box detector. Just counted patterns, reported as measured, including the two findings that surprised us.
AI text is over six times as likely to contain a stock phrase
11.3% of AI passages contained at least one stock phrase ("it's important to note", "in conclusion") versus 1.8% of human passages. Per word, stock phrases appeared at 0.64 per 1,000 words in AI text against 0.13 in human text. This was the single strongest signal we measured.
"Additionally" is the loudest single word
"Additionally" appeared 341 times per million words in AI answers versus 59 in human answers, a 5.6× gap. "Navigate" showed a 3.8× gap. Not every suspect word earned its reputation, though: "testament" appeared at nearly identical rates in both groups.
AI sentence rhythm really is flatter
Human passages varied their sentence lengths with a standard deviation of 9.3 words; AI passages measured 7.3, about 22% less variation. Humans mix short punches with long, winding sentences. The model settles into a cadence. This burstiness gap is why sentence rhythm is the heaviest-weighted signal in our style engine.
The em-dash tell is generational, and the data made us say so
The famous "AI loves em-dashes" tell was inverted in this corpus: 2023-era ChatGPT used em-dashes at 0.001 per 100 words, roughly 50× less than the humans (0.056). The em-dash habit belongs to later models. We publish this even though our own checker flags em-dash density, because the honest reading is that tells are model-generational and need re-measuring as models change.
Contractions barely separated the groups
Human answers contracted 41.0% of the time, AI answers 38.2%. In casual Q&A register, ChatGPT contracts nearly as much as people do. Contraction rate is a fingerprint of a specific writer, not a reliable human-versus-AI divider, which is exactly how our style engine uses it.
The numbers
| Measure | Human (n=4,000) | ChatGPT (n=4,000) |
|---|---|---|
| Total words analyzed | 675,404 | 741,334 |
| Passages with ≥1 stock AI phrase | 1.8% | 11.3% |
| Stock phrases per 1,000 words | 0.13 | 0.64 |
| AI-tell words per 1,000 words | 0.25 | 0.46 |
| Sentence-length std deviation (words) | 9.3 | 7.3 |
| Em-dashes per 100 words | 0.056 | 0.001 |
| Contraction rate | 41.0% | 38.2% |
Method, honestly
Data. The HC3 corpus (Hello-SimpleAI, 2023): paired human and ChatGPT answers to identical questions. We used the reddit_eli5 and open_qa splits, sampling 4,000 passages per group at 80 to 400 words each.
Analysis. The same deterministic rules that power our free AI-tell checker: a fixed vocabulary of tell words, a fixed library of stock phrases, and computed stylometrics (sentence-length distribution, punctuation densities, contraction rate). Counting, not classifying.
Limits, stated plainly. The AI text is 2023-era ChatGPT; newer models have different tells (the em-dash finding shows exactly that). The humans are Reddit and forum answerers writing casually in English. None of this makes a detector: 3.5% of human passages tripped a tell word, and plenty of AI passages tripped nothing. Patterns shift probability; they do not prove authorship.
Reproducibility. Every rule we counted with runs in your browser on the checker page. Take any HC3 passage and verify our flags yourself. Our scoring method is documented in how the style match score works.
Citing this: “Shlokah Research, What Statistically Gives AI Writing Away, July 2026, shlokah.com/research.” Corpus credit: Guo et al., HC3 (Hello-SimpleAI), CC-BY-SA.
Check your own text against these patterns
The same rules, free and unlimited, running entirely in your browser. Nothing is uploaded.
Run the AI-tell checkerAlso in this series: Measured #1: the DIY voice-guidelines method, tested