The 5 Biggest "Tells" That Something Was Written By AI
Writers are reading with new eyes. A clean paragraph now triggers a second question: who wrote this-someone with a pulse, or a machine with great manners?
The concern isn't just academic. Machine-written text carries patterns that can blur intent, trust, and voice across briefs, newsletters, and newsrooms. Research on stylometry keeps finding the same signals: predictable sentence shapes, steady cadence, and function-word habits that look more algorithmic than alive.
These signals are shrinking as models evolve-remember the em dash obsession? Still, the gap is big enough to guide editors. None of the traits below prove authorship; they only push probability. Stack a few together, and the signal gets loud.
1) Negative parallelism and oversimplified contrast
AI leans hard on neat pivots: "It's not X, it's Y." Or "not just X-also Y." It reads tidy, suggests insight, and often says very little. Large datasets of model outputs show these mirrored clauses popping up far more than in human drafts.
- Spot it: Repeated hinge phrases that compress a complex idea into a binary flip.
- Fix it: Replace the flip with one concrete example, a number, or a short story. Specifics beat symmetry.
2) Over-neat structure and a suspiciously even rhythm
Machine text glides. Every paragraph lands on rails. Transitions are frictionless. Sentence lengths hover in the same range. Human writing-yes, even polished work-wanders a bit, takes a sharp turn, then sprints.
- Spot it: Paragraphs that all look alike. Transitions that sound like a handbook: "First… Next… Finally…"
- Fix it: Read aloud. Shorten one sentence to five words. Let the next run long. Drop a transition and make the thought jump.
3) Smoothed-out emotion and extra-polite hedging
Models aim to please. They cushion with phrases like "It's understandable that…" or close with tidy recaps. The tone hovers at friendly-customer-support, even when the subject calls for bite or surprise. Studies show model sentiment swings less than human tone.
- Spot it: Apologies, softeners, and round-edged conclusions that pat the reader on the head.
- Fix it: Say the thing. Cut two hedges per paragraph. Swap "It might be helpful to consider" for a direct verb and an example.
4) Vague abstractions and "safe" vocabulary
When ideas get thin, machines reach for fog: "ecosystem," "framework," "dynamic," "solution," "synergy." You've seen it. Research keeps finding lower lexical variety and heavier noun stacks in AI drafts. Also, the buzzwords shift over time-so word lists age fast; structure is the steadier tell.
- Spot it: Dense clusters of abstract nouns with no names, numbers, or scenes.
- Fix it: Name a person, a place, a date, or a metric. Replace three abstractions with one concrete detail.
5) Balanced clauses and conspicuously careful phrasing
"While X is true, Y is also important." "Whether you're a beginner or an expert…" These constructions feel safe because they avoid commitment. Stylometric work shows models overuse these function-word patterns compared to human baselines.
- Spot it: Sentences that bend over backwards to be fair, everywhere.
- Fix it: Pick a side. Write one sharp line that would make a colleague push back. Then earn it with evidence.
How to edit a draft that feels AI-ish
- Break the rhythm: Aim for obvious variation-one short sentence, one long. Then a mid-length one.
- Cut the cotton: Delete hedges, courtesy fillers, and wrap-up phrases. Keep the claim; lose the fluff.
- Trade fog for facts: Replace abstractions with names, sources, screenshots, or numbers.
- Add friction: Insert a counterexample or a sharp aside. Humans contradict themselves; let a little mess live.
- Show, then say: Lead with a scene, result, or line of dialogue. Follow with the point in plain English.
- Use the "one change" rule: Revise at least 30% of any AI-assisted draft with your own examples or experience.
If you use AI in your process
- Ask for raw notes, not final prose: Bullets, sources, edge cases. You do the writing.
- Interrogate claims: "Where did this come from?" "What would contradict it?"
- Force specificity: Require three concrete examples before any summary.
Bottom line: detection will stay probabilistic, and models will keep smoothing their seams. As a writer, your edge is voice, judgment, and proof. Keep those front and center, and the text will read unmistakably human.
Want structured practice? Explore practical AI courses for writers and editors at Complete AI Training - Courses by Job or sharpen prompting with Prompt Engineering resources.
Your membership also unlocks:
