JFly.Ai blog article on making AI-assisted writing read human again: measured contraction rates, hedge cutting, teaching AI your voice from real sent mail, and testing your own writing checkers.

JFly.Ai Book a call

AI & Writing

How to Stop Sounding Like AI Wrote It

My AI turned into a terrible writer. I audited every writing instruction it carries, measured my own emails against its drafts, and found that three fixes did almost all the work. Here they are with the numbers.

The whole thing in fifteen seconds

  • The loudest machine tell is missing contractions, not fancy words.
  • Hedge words are a third of your length and none of your meaning.
  • Teach AI your voice from mail you actually sent, never from its own drafts.
  • Specific, named, real detail is the one thing polish can't fake.

01

The tell isn't the vocabulary

Almost none of my writing rules ever loaded. I found 34,973 words of carefully written guidance that had fired about 7 times in six weeks, while the 479 words that did load on every single message contradicted them. It wasn't a bad writer. It was obeying the wrong file.

So I stopped guessing and measured. I ran 14 human articles against 21 machine ones. The biggest difference wasn't word choice. It was contractions: the humans used 19.2 per thousand words, the machine 4.1. Machines write like a policy document.

Measured, not vibes

Contractions per thousand words

Human writers 19.2
Machine drafts 4.1
Where these numbers come from: my own corpus test, August 2026. 14 human-written articles measured against 21 machine-written ones.

The fix costs nothing: read your last email out loud. Every "do not" becomes "don't," every "it is" becomes "it's," everywhere except contracts. If you wouldn't say the long form across a table, don't type it.

02

Cut the hedges

I cut a client document from 4,989 words to 2,267 across seven rounds and never deleted a fact. The last five hundred words were pure hedging. Every "we would" and "worth checking" is a sentence apologizing for itself.

Before / after

The same sentence with its apology removed

Hedged

We'd use Microsoft

What we would build

Worth checking first

Committed

Microsoft

What we build

Check that line first

Real edits from one client document that went from 4,989 words to 2,267 with zero facts removed.

Search your draft for would, could, and worth, then at each hit either commit to the claim or cut it, because a document stripped of its apologies says the same thing twice as fast.

03

Teach it from what you actually sent

You can't teach a machine your voice from things the machine wrote. The loop feeds its accent straight back.

This week I mined ten years of my own sent email across three accounts, and the best examples were ones no style guide would produce: a decline with typos left in, a lease note owning a missed deposit in two words, a thank-you with the letters stretched. Mail sent before AI existed is guaranteed you. That older mail beat every polished recent draft.

Before I let a filter judge my writing, I ran it on 54 emails I'd actually sent, and my real phrases passed while only the machine-drafted lines got flagged. If a checker flags your own voice, the checker is wrong, not you.

You can't teach a machine your voice from things the machine wrote. Only from what you actually sent.

The rule that rebuilt my whole voice file

04

Specificity is the last mile

Fixing sentences took my writing from 99% flagged as machine-made down to about 45%. It never reached zero. What separates human writing isn't craft. It's that a person went and found something out: a real number, a named app, a date, a thing that happened.

Same draft, same detector

Flagged as machine-made

Before the fixes 99%
After the fixes ~45%
The gap that remains is evidence, not style. The detector rewarded named, specific, real detail more than any sentence-level polish.

Polish is cheap now and evidence isn't: one found fact beats three well-made sentences.

05

Test whatever grades you

I fed a popular writing tool prose I knew a person wrote, and it flagged the human harder than the machine, 22.0 versus 10.4 per thousand words. Before you trust anything that grades your work, feed it something you already know the answer to. The graders need grading too.

The cleanup pass

Six moves, one afternoon

  • Read your last three sent emails out loud; collapse every long form you'd never say.
  • Search one live document for would, could, and worth; commit or cut at every hit.
  • Pull 20 emails you sent before AI existed into one file. That's your voice sample.
  • Build a short banned list from words you never say out loud, and enforce it with a checker, not willpower.
  • Put one real number or named thing in every claim.
  • Feed your grader known-human writing before you believe a word it says.

Questions we get

Does this apply if AI drafts most of my email?
More, not less: the drafts inherit whatever voice sample you gave the machine. Fix the sample and every draft downstream improves.
Which punctuation gives AI away?
The long dash is the famous one, but the measured tell is bigger: uncontracted prose. Fix that first.
How many examples does an AI need to learn your voice?
Fewer than you think. I cap mine at six per type of writing and drop the weakest when a better one lands.

Let's Build Your AiOS.

Your voice is one layer of it. Book a call and we wire your writing rules, your banned list, and your real voice into the system that drafts for you.

Book a call

Blueprint → Build → Partner · Denver + remote