AI Writing Detection
10 minAI Writing Detection
When a language model writes prose, it leaves fingerprints. These patterns appear 5 to 20 times more often in machine-generated text than in human writing. Recognising them is the first step to editing them out. Ignoring them means shipping copy that every reader who has used a chatbot will instantly tag as generated.
Three categories of tells: vocabulary, structure, and chatbot artifacts. Each has different severity, and fixing them follows a clear priority order.
Tier 1 words: replace on sight
The Tier 1 word list is the small set of vocabulary that appears so disproportionately in AI prose that any one of them is a tell. These words exist because RLHF training promoted them ("robust," "comprehensive") or because the model treats them as upgrades to plain alternatives ("leverage" over "use").
| AI word | Replace with | |---|---| | delve / delve into | explore, dig into, look at | | landscape (metaphor) | field, space, industry | | robust | strong, reliable, solid | | comprehensive | thorough, complete, full | | leverage (verb) | use | | seamless / seamlessly | smooth, easy, without friction | | cutting-edge | latest, newest, advanced | | pivotal | important, key, critical | | utilize / utilise | use | | holistic | complete, full, whole | | actionable | practical, useful, concrete | | impactful | effective, significant | | empower | let, help, enable | | streamline | simplify, speed up | | harness | use | | foster | build, encourage, support | | in order to | to | | due to the fact that | because | | serve as | is | | learnings | lessons, findings, takeaways | | best practices | what works, proven methods | | synergy | (describe the combined effect) | | thought leadership | expertise | | deep dive | look at, examine, explore | | game-changer | (describe what changed) |
The plain alternative is almost always better. The most reliable improvement to LLM prose is purely mechanical: replace each Tier 1 word with the plain version. The result reads more direct, more confident, and more human.
When the Tier 1 word stays: direct quotes from a source (don't edit other people's words), genuine terms of art ("robust" in software security has a specific technical meaning), and cases where the word is genuinely the right one after deliberation.
Structural patterns
Beyond vocabulary, AI text has structural tics that trained readers spot instantly.
Em-dash overuse. Hard cap: one per 1,000 words. Zero is better. AI uses the em dash because it's syntactically flexible. Human writers use it sparingly. Replace with a comma, a full stop, parentheses, or two sentences.
Bold overuse. Strip bold from most phrases. Maximum one bolded phrase per major section, ideally none. If something needs to stand out, restructure the sentence to lead with it.
The Rule of Three. Every AI list has exactly three items. Sometimes the right number is two, four, or seven. Three is a default, not a decision.
Uniform paragraph length. Every paragraph is three sentences. Real prose has rhythm: one-sentence paragraphs, six-sentence paragraphs, whatever the content needs.
Synonym cycling. "Developers... engineers... practitioners... builders" in one paragraph to avoid repetition. Human writers repeat the clearest word.
"It's not X, it's Y." Once is fine. AI uses this construction every other paragraph. Vary the rhythm.
Copula avoidance. AI substitutes "serves as," "features," "boasts," "presents" for "is" and "has." Default to "is" and "has" unless a specific verb adds meaning.
Chatbot artifacts: P0 credibility killers
Chatbot artifacts are the conversational openers, closers, and acknowledgements that a chat-tuned model produces by default. They're appropriate in chat. They're an immediate credibility killer in published prose.
Sycophantic openers: "Great question!" / "Excellent point!" / "That's a fascinating area." In chat, politeness. In published copy, desperate.
Helpful-assistant closers: "I hope this helps!" / "Feel free to reach out!" / "Happy to clarify!" These belong in customer-service emails, not marketing copy.
"Let's + verb" transitions: "Let's dive in!" / "Let's take a look at..." / "Let's explore..." Any "Let's + verb" functioning as a transition rather than a genuine invitation. Start with the point instead.
Chain-of-thought leakage: "Breaking this down..." / "To approach this systematically..." / "Step 1:" These betray the reasoning machine underneath. State the conclusion, then the evidence.
Cutoff disclaimers: "As of my last update..." / "Based on information available to me..." Never appropriate in published copy.
A chatbot artifact in a hero section is worse than a Tier 1 word. It tags the entire site as machine-written. One "Let's dive in!" on an article opening cannot survive to publication.
The severity ladder
Fix in this order:
P0 (credibility killers): Chatbot artifacts, vague attributions ("experts believe," "studies show" without citing the expert or study), significance inflation ("marking a pivotal moment in the evolution of...").
P1 (obvious AI smell): Tier 1 word violations, "Let's" transition openers, formulaic openings, bold overuse.
P2 (polish): Uniform paragraph length, copula avoidance, synonym cycling. Fix when time allows. These are subtle enough that only trained readers will catch them.
A copy pass that removes nothing else but eliminates chatbot artifacts is already a significant improvement. Start with P0.