The voice problem used to be a vibe. Now it's a checklist. Readers, judges, and critics have started naming the exact tics that make a paragraph feel machine-touched, and if you use tools like YouWrite's refine pass, you owe your draft a serious read before it ships.
This isn't about beating detectors. Detectors are unreliable and getting worse. It's about not sounding like everyone else who opened the same tab this morning.
The complaints are getting specific
In May 2025, the Commonwealth Short Story Prize disqualified entries its judges suspected were AI-generated. Coverage in The Guardian quoted judges describing a strange sameness: technically fluent prose that felt oddly weightless, with emotional beats that stated feelings instead of embodying them. Around the same time, Slate ran a piece by Laura Miller on book critics learning to spot AI in review copy, focusing on transitional filler, confident but content-free summary sentences, and a peculiar allergy to specificity.
What used to be reader intuition ("something's off") is now a growing bill of particulars. Good. Vague complaints don't help you revise. Specific ones do.
The fingerprints
Some of these appear in human writing too. That's fine. One or two won't dent your voice. A cluster of them, paragraph after paragraph, is the tell.
1. Punctuation as smoothing agent
Em dashes and en dashes are the most notorious. AI models reach for them constantly because they paper over transitions a careful writer would earn with a period or a comma. If your draft has three em dashes on a page and you didn't type any of them, that's a signal, not a stylistic choice.
Semicolons behave similarly when overused. Watch for a rhythm where every long sentence hinges on the same joint.
2. Hollow intensifiers and meta-phrases
"It is worth noting." "Importantly." "Notably." "That said." "At its core." These phrases pretend to add emphasis while adding nothing. Delete them and the sentence almost always improves. If it doesn't, the sentence itself was the problem.
Related tell: the pre-announcement. "Let's consider three reasons." Just give the reasons.
3. Frictionless transitions
AI loves a bridge sentence that restates the previous paragraph in gentler terms before pivoting. "This raises a broader question about how we think about X." That's a stage direction, not a thought. Cut it and see if the two paragraphs still talk to each other. Usually they do.
4. Confident vagueness in emotional beats
This is the one the Commonwealth judges flagged. The prose asserts an emotion instead of rendering it. "She felt a complex mixture of grief and relief." Compare with the concrete version: she keeps folding the same dish towel. One tells, one shows, and the machine prefers the first because the first is safer and requires no observation of the world.
5. The tri-colon reflex
Everything comes in threes. Rhythms, examples, adjectives. "Clear, concise, and compelling." Once per essay is fine. Every paragraph is a fingerprint.
6. Abstract nouns doing verbs' work
"The implementation of the strategy resulted in an improvement in engagement." Rewrite: "After we changed the headline, clicks doubled." AI-assisted drafts nominalize compulsively because abstract nouns are safer than concrete verbs. They also weigh twice as much and say half as much.
7. The false-balance sentence
"While X has its merits, it is also important to acknowledge Y." This construction shows up when the model is hedging to sound judicious. Real judiciousness picks a side and defends it, or names the specific tradeoff.
8. Setting the scene without setting anything
"In an era of rapid technological change..." "As the world becomes increasingly..." Throat-clearing at industrial scale. Start with the sentence that comes after.
The audit, sentence by sentence
Open your draft. For each paragraph, run this pass:
- Highlight every em dash, en dash, and semicolon. Keep the ones you'd defend in a workshop. Convert the rest.
- Search for: worth noting, importantly, notably, that said, at its core, in essence, ultimately, crucial, vital, delve, leverage, navigate, unlock, tapestry, landscape. Delete or replace with something specific.
- Find every sentence that names an emotion with an abstract noun (grief, joy, unease, tension). Ask what a camera would see. Replace at least half with the physical detail.
- Read the first sentence of each paragraph aloud. If it summarizes the previous paragraph, cut it.
- Count your rule-of-three lists. Keep one. Break the others into asymmetric pairs or fours.
- Look for sentences of similar length in a row. Break the pattern. A three-word sentence after a twenty-two-word one changes the pulse.
- Find every claim that could be attached to a real name, number, or date. Attach it. If you can't, ask whether the claim is doing any work.
The audit takes about twenty minutes on a thousand-word piece. It's slower than generating another draft. That's the point.
Where this critique is fair to AI, and where it isn't
Humans write bad transitions and hollow intensifiers too. Corporate blog posts have looked like this since long before GPT-2. What's new isn't the existence of the patterns, it's the density. A ChatGPT draft can pack all eight fingerprints into a single page, which almost never happens in unedited human writing.
Honest note about YouWrite: our refine tool is good at tightening structure and cutting flab, but it will not invent the concrete detail your third paragraph is missing. No refiner can. That has to come from you or from your notes. Hand a refiner a draft with no observed detail, and you get a smoother draft with no observed detail. The specificity is your job.
What to do with the audit results
Don't try to score a zero. A piece with two em dashes and one "notably" is not compromised. A piece where every emotional beat is abstract, every transition is a bridge sentence, and every list has three items has a voice problem regardless of who or what wrote it.
Run refine on the sentences. Keep the noticing for yourself.