AI writing tells, and what each one means
People arguing about AI writing mostly cannot name what they are looking at. These are the names, with an example of each, so the argument has words in it.
Under every definition is what happened when the pattern was counted across 110,305 words: 4 models and 2 sets of human writing from before ChatGPT existed. Most of these patterns are cited far more often than they occur. That is worth knowing before you accuse anybody of anything.
The machine-readable version is rules.json, generated from the same source as the checker, so this page cannot drift away from the tool.
- AI-favorite word
- Chatbot artifact
- Contrastive negation
- Editorializing aside
- Em-dash density
- Fake-drama scaffolding
- Fragment triplet
- Hedged balance
- Inflated symbolism
- Negative parallelism
- Rule of three
- Self-narration
- Stock opener
- Summary phrase
- Trailing '-ing' analysis
- Vague attribution
AI-favorite word Loudest
A word that machine drafts reach for far more often than people do. Delve, leverage, showcase, meticulous, pivotal, seamless.
Example We leveraged the data to showcase pivotal insights.
Instead We used the data to show what mattered.
Highest in Gemini Flash, Aug 2026 at 1.01, but the interval overlaps human writing. Not separation.
| Corpus | Rate | Interval |
|---|---|---|
| Claude Opus 5, Aug 2026 | 0.09 | 0.00–0.23 |
| GPT-5.5, Aug 2026 | 0.11 | 0.00–0.29 |
| ChatGPT, Aug 2026 | 0.09 | 0.00–0.29 |
| Gemini Flash, Aug 2026 | 1.01 | 0.54–1.55 |
| Human, Medium articles | 0.33 | 0.13–0.57 |
| Human, newsletters | 0.57 | 0.15–1.22 |
Chatbot artifact Legacy
Residue from a chat window that got copied along with the text.
Example Great question! Let me know if you'd like me to expand on any of this.
Instead Delete it.
Fired zero times across 110,305 words. Widely cited, not measurable here.
| Corpus | Rate | Interval |
|---|---|---|
| Claude Opus 5, Aug 2026 | 0.00 | 0.00–0.00 |
| GPT-5.5, Aug 2026 | 0.00 | 0.00–0.00 |
| ChatGPT, Aug 2026 | 0.00 | 0.00–0.00 |
| Gemini Flash, Aug 2026 | 0.00 | 0.00–0.00 |
| Human, Medium articles | 0.00 | 0.00–0.00 |
| Human, newsletters | 0.00 | 0.00–0.00 |
Contrastive negation Legacy
The same move with the intensifier removed, usually split across two sentences. Defining a thing by what it isn't.
Example This isn't a verdict. It's a place to look.
Instead This is a place to look.
Elevated in Claude Opus 5, Aug 2026 (0.92) and GPT-5.5, Aug 2026 (1.90), against 0.15 at most in human writing. Shared by more than one model, so it is not a fingerprint.
| Corpus | Rate | Interval |
|---|---|---|
| Claude Opus 5, Aug 2026 | 0.92 | 0.46–1.52 |
| GPT-5.5, Aug 2026 | 1.90 | 0.96–2.71 |
| ChatGPT, Aug 2026 | 0.46 | 0.09–0.98 |
| Gemini Flash, Aug 2026 | 0.14 | 0.00–0.30 |
| Human, Medium articles | 0.15 | 0.03–0.30 |
| Human, newsletters | 0.10 | 0.00–0.33 |
Editorializing aside Structural
A sentence that tells the reader a fact is important rather than giving them the fact.
Example It's important to note that the deadline has moved.
Instead The deadline moved to Friday.
Too rare to characterize: 1 hit in 110,305 words across every corpus.
| Corpus | Rate | Interval |
|---|---|---|
| Claude Opus 5, Aug 2026 | 0.00 | 0.00–0.00 |
| GPT-5.5, Aug 2026 | 0.00 | 0.00–0.00 |
| ChatGPT, Aug 2026 | 0.00 | 0.00–0.00 |
| Gemini Flash, Aug 2026 | 0.00 | 0.00–0.00 |
| Human, Medium articles | 0.04 | 0.00–0.12 |
| Human, newsletters | 0.00 | 0.00–0.00 |
Em-dash density Structural
How often the long dash appears. People use it; some models use it several times more often than people do.
Example The result was clear — and it changed everything — or so it seemed.
Instead A comma, a period, or a pair of parentheses will usually do the same work.
A Claude Opus 5, Aug 2026 habit. 11.05 per 1,000 words, and its interval clears every other model and both human sets.
| Corpus | Rate | Interval |
|---|---|---|
| Claude Opus 5, Aug 2026 | 11.05 | 9.59–12.53 |
| GPT-5.5, Aug 2026 | 1.63 | 0.78–2.75 |
| ChatGPT, Aug 2026 | 3.30 | 1.85–4.84 |
| Gemini Flash, Aug 2026 | 2.70 | 2.00–3.42 |
| Human, Medium articles | 3.83 | 2.67–5.12 |
| Human, newsletters | 1.05 | 0.22–2.24 |
Fake-drama scaffolding Legacy
Canned suspense wrapped around an ordinary point.
Example Here's the part nobody tells you. Let that sink in.
Instead State the point.
Too rare to characterize: 1 hit in 110,305 words across every corpus.
| Corpus | Rate | Interval |
|---|---|---|
| Claude Opus 5, Aug 2026 | 0.05 | 0.00–0.14 |
| GPT-5.5, Aug 2026 | 0.00 | 0.00–0.00 |
| ChatGPT, Aug 2026 | 0.00 | 0.00–0.00 |
| Gemini Flash, Aug 2026 | 0.00 | 0.00–0.00 |
| Human, Medium articles | 0.00 | 0.00–0.00 |
| Human, newsletters | 0.00 | 0.00–0.00 |
Fragment triplet Loudest
Three clipped sentence fragments in a row, borrowing cadence without adding information.
Example Faster. Cleaner. Simpler.
Instead One sentence that says what changed.
Too rare to characterize: 1 hit in 110,305 words across every corpus.
| Corpus | Rate | Interval |
|---|---|---|
| Claude Opus 5, Aug 2026 | 0.00 | 0.00–0.00 |
| GPT-5.5, Aug 2026 | 0.05 | 0.00–0.18 |
| ChatGPT, Aug 2026 | 0.00 | 0.00–0.00 |
| Gemini Flash, Aug 2026 | 0.00 | 0.00–0.00 |
| Human, Medium articles | 0.00 | 0.00–0.00 |
| Human, newsletters | 0.00 | 0.00–0.00 |
Hedged balance Loudest
Weighing both sides at length without ever landing on one.
Example While there are challenges, there are also opportunities.
Instead Say which one you think wins, or name the one thing that is genuinely uncertain.
Too rare to characterize: 2 hits in 110,305 words across every corpus.
| Corpus | Rate | Interval |
|---|---|---|
| Claude Opus 5, Aug 2026 | 0.00 | 0.00–0.00 |
| GPT-5.5, Aug 2026 | 0.00 | 0.00–0.00 |
| ChatGPT, Aug 2026 | 0.09 | 0.00–0.28 |
| Gemini Flash, Aug 2026 | 0.00 | 0.00–0.00 |
| Human, Medium articles | 0.04 | 0.00–0.12 |
| Human, newsletters | 0.00 | 0.00–0.00 |
Inflated symbolism Structural
Treating an ordinary fact as though it stands for something larger.
Example The launch stands as a testament to the team's dedication.
Instead The team shipped it in six weeks.
Fired zero times across 110,305 words. Widely cited, not measurable here.
| Corpus | Rate | Interval |
|---|---|---|
| Claude Opus 5, Aug 2026 | 0.00 | 0.00–0.00 |
| GPT-5.5, Aug 2026 | 0.00 | 0.00–0.00 |
| ChatGPT, Aug 2026 | 0.00 | 0.00–0.00 |
| Gemini Flash, Aug 2026 | 0.00 | 0.00–0.00 |
| Human, Medium articles | 0.00 | 0.00–0.00 |
| Human, newsletters | 0.00 | 0.00–0.00 |
Negative parallelism Loudest
Saying what something is not, immediately before saying what it is, with an intensifier in the middle.
Example It's not just a checker, it's a way of thinking about writing.
Instead It's a checker.
Highest in ChatGPT, Aug 2026 at 0.55, but the interval overlaps human writing. Not separation.
| Corpus | Rate | Interval |
|---|---|---|
| Claude Opus 5, Aug 2026 | 0.05 | 0.00–0.14 |
| GPT-5.5, Aug 2026 | 0.22 | 0.05–0.45 |
| ChatGPT, Aug 2026 | 0.55 | 0.11–1.10 |
| Gemini Flash, Aug 2026 | 0.37 | 0.09–0.72 |
| Human, Medium articles | 0.22 | 0.05–0.43 |
| Human, newsletters | 0.10 | 0.00–0.32 |
Rule of three Loudest
Three near-synonyms in a row, used for rhythm rather than meaning.
Example An innovative, transformative, and groundbreaking approach.
Instead Keep the one word that is true.
Highest in ChatGPT, Aug 2026 at 0.46, but the interval overlaps human writing. Not separation.
| Corpus | Rate | Interval |
|---|---|---|
| Claude Opus 5, Aug 2026 | 0.09 | 0.00–0.23 |
| GPT-5.5, Aug 2026 | 0.33 | 0.10–0.67 |
| ChatGPT, Aug 2026 | 0.46 | 0.10–0.92 |
| Gemini Flash, Aug 2026 | 0.18 | 0.04–0.37 |
| Human, Medium articles | 0.11 | 0.00–0.25 |
| Human, newsletters | 0.10 | 0.00–0.30 |
Self-narration Instruction file only
Writing about the writing instead of about the subject. Rating your own confidence, announcing your honesty, apologizing for an earlier draft, or telling the reader how much something matters. Not in the checker, because it has no fixed shape to match.
Example Worth saying, because it is the more interesting half.
Instead Cut it. If the reader loses no fact about the subject, it was self-narration.
Named and defined, deliberately not in the checker. Its forms are unbounded, so no pattern can match them.
Stock opener Legacy
An opening line that could sit on top of any article on any subject.
Example In today's fast-paced world, communication matters more than ever.
Instead Open with the thing you actually found.
Too rare to characterize: 1 hit in 110,305 words across every corpus.
| Corpus | Rate | Interval |
|---|---|---|
| Claude Opus 5, Aug 2026 | 0.00 | 0.00–0.00 |
| GPT-5.5, Aug 2026 | 0.05 | 0.00–0.16 |
| ChatGPT, Aug 2026 | 0.00 | 0.00–0.00 |
| Gemini Flash, Aug 2026 | 0.00 | 0.00–0.00 |
| Human, Medium articles | 0.00 | 0.00–0.00 |
| Human, newsletters | 0.00 | 0.00–0.00 |
Summary phrase Structural
A closing that restates what the reader has just read instead of ending on the last new thing.
Example In conclusion, the data shows a clear trend.
Instead End on the trend itself.
Too rare to characterize: 6 hits in 110,305 words across every corpus.
| Corpus | Rate | Interval |
|---|---|---|
| Claude Opus 5, Aug 2026 | 0.00 | 0.00–0.00 |
| GPT-5.5, Aug 2026 | 0.05 | 0.00–0.21 |
| ChatGPT, Aug 2026 | 0.09 | 0.00–0.33 |
| Gemini Flash, Aug 2026 | 0.05 | 0.00–0.15 |
| Human, Medium articles | 0.11 | 0.00–0.28 |
| Human, newsletters | 0.00 | 0.00–0.00 |
Trailing '-ing' analysis Structural
An -ing clause bolted onto the end of a sentence to make a plain fact sound analyzed.
Example Revenue rose 12 percent, underscoring the strength of the strategy.
Instead Revenue rose 12 percent.
Fired zero times across 110,305 words. Widely cited, not measurable here.
| Corpus | Rate | Interval |
|---|---|---|
| Claude Opus 5, Aug 2026 | 0.00 | 0.00–0.00 |
| GPT-5.5, Aug 2026 | 0.00 | 0.00–0.00 |
| ChatGPT, Aug 2026 | 0.00 | 0.00–0.00 |
| Gemini Flash, Aug 2026 | 0.00 | 0.00–0.00 |
| Human, Medium articles | 0.00 | 0.00–0.00 |
| Human, newsletters | 0.00 | 0.00–0.00 |
Vague attribution Structural
Crediting a claim to an unnamed authority so it sounds sourced without being sourced.
Example Experts say remote work is here to stay.
Instead Name the expert, or make the claim yourself.
Fired zero times across 110,305 words. Widely cited, not measurable here.
| Corpus | Rate | Interval |
|---|---|---|
| Claude Opus 5, Aug 2026 | 0.00 | 0.00–0.00 |
| GPT-5.5, Aug 2026 | 0.00 | 0.00–0.00 |
| ChatGPT, Aug 2026 | 0.00 | 0.00–0.00 |
| Gemini Flash, Aug 2026 | 0.00 | 0.00–0.00 |
| Human, Medium articles | 0.00 | 0.00–0.00 |
| Human, newsletters | 0.00 | 0.00–0.00 |
How this was measured. Rates are hits per 1,000 words, with a 95% confidence interval bootstrapped over whole documents rather than sentences. A pattern is called one model's habit only when its interval clears every other model and both human sets. Anything under 0.20 per 1,000 words is reported as too rare to characterize, because a single hit in twenty thousand words is noise. Corpora, generator, and measurement scripts are in study/. This page is generated; the definitions live in glossary-definitions.json.