nli33 AI vs Human Writing Patterns
Overview
A reference list of surface patterns correlated with AI-generated vs human-written text. Use it to identify patterns present in a piece of text. No pattern alone proves authorship — density and combination matter more than any single instance. This skill does not say whether AI-sounding or human-sounding text is better; that judgment belongs to whoever is using the skill.
AI-leaning patterns
Negative parallelism — "not just X but Y," "it's not X, it's Y," "Y rather than X," "Y instead of X," "no X, no Y, just Z."
Rule of three — near-every list, description, or takeaway comes in exactly three parts (three adjectives, three benefits, three clauses).
Em dash overuse — em dashes (spaced, —) standing in for commas, colons, or parentheses, often for false emphasis or "punchy" sales-style cadence.
Buzzword/AI vocabulary — delve, tapestry, landscape, realm, embark, pivotal, crucial, intricate, meticulous, multifaceted, comprehensive, robust, leverage, underscore, showcase, testament, boasts, stands as, serves as, "it's important to note," furthermore. (Word list drifts by model generation — density of these terms is the signal, not any one word.)
"By [gerund] X, Y" constructions — opening a sentence, especially a conclusion, with "By doing X, we can achieve Y."
False hot-take framing — "But here's the truth:", "Here's the thing:", "The reality is..." — a manufactured contrarian turn with no real stakes, common in LinkedIn/corporate copy.
Manufactured friendliness — upbeat, eager-to-please tone; exclamation points and affirmations ("Great question!", "I'd be happy to help!") applied uniformly regardless of content.
Uniform sentence rhythm — sentences cluster around similar length and structure; few short fragments or run-ons breaking the pattern.
Hedging modals — frequent "might," "could," "may" softening claims without a clear reason.
Assertive close, low doubt markers — confident, tidy conclusions with little expressed uncertainty or self-correction, even on contestable claims.
Elevated formality/structure — consistent topic-sentence-then-support paragraphing, frequent transition words (furthermore, moreover, additionally), heavy subheadings/bullets even in casual contexts.
High content-word density, low function-word variety — fewer articles/prepositions/conjunctions relative to content words than typical human prose.
Human-leaning patterns
Variable sentence length — mix of short fragments, long run-ons, and everything between.
Irregular structure — asides, tangents, self-interruptions, afterthoughts, unresolved threads.
Idiosyncratic vocabulary — slang, in-jokes, domain-specific shorthand, inconsistent register within one piece.
Visible emotion, including negative — frustration, sarcasm, disgust, boredom — flatter/rarer in AI text, which skews positive.
Personal/concrete references — specific names, dates, inside context, first-person anecdote rather than generic examples.
Imperfections — typos, informal punctuation, sentence fragments, inconsistent capitalization, self-correction mid-thought.
Understatement or flat close — endings that trail off or land plainly rather than a tidy summary bow.
Using this list
When asked to check text: scan for which patterns appear, note density (a single em dash or "crucial" proves nothing; five AI-leaning patterns clustered in one paragraph is a signal), and report findings neutrally. Do not rewrite or "fix" the text unless separately asked.