How to Make AI Writing Sound Human: Say Something That Could Be Wrong

2026-09-07 · The Prowir Team

Every few months the internet finds a new way to accuse people of using AI, and right now it is punctuation. Somebody spots an em dash, declares the case closed, and a competent writer spends an afternoon defending their own commas. Writers have started stripping dashes out of their drafts preemptively. Model makers have started tuning the character down. That is a lot of energy spent on the wrong signal. If you want to know how to make AI writing sound human, punctuation is the last place to look. Readers are not counting dashes. They are noticing that nothing in the piece could possibly be wrong.

Punctuation is a terrible detector

The em dash predates every model by a few centuries. It is a normal tool used by people who write a lot, which is exactly why it shows up in text trained on people who write a lot. Treating it as evidence is the same error automated detectors keep making: research on AI detection tools found they flag writing by non-native English speakers at far higher rates than native speakers, because plainer, more regular sentence patterns look statistically machine-like. The tool is not detecting machines. It is detecting a surface habit and calling it a confession.

There is a simpler reason to drop the dash hunt. Any tell that a four-word prompt can remove is not a tell. It is a fashion. The models have already moved; the accusation industry will need a new punctuation mark by spring. Meanwhile the thing that actually gives away machine-drafted work has not changed at all.

The real tell is the absence of risk

Read enough AI drafts and the uniform quality is not the typography. It is the refusal to commit. Every claim arrives pre-hedged. Every recommendation ends in "it depends on your organization." Both sides get their fair paragraph, nobody gets blamed, and the piece closes on a sentence that could be pasted into any article on any topic without anyone noticing.

That is not an accident of style, it is the training objective showing through. A model optimized to be helpful and agreeable is optimized to avoid being wrong, and prose that cannot be wrong cannot be interesting. There is nothing to grab. A reader finishes it, agrees with all of it, and remembers none of it.

So use the test that actually works: could a competent person in your field read this and disagree with you? Not misunderstand you — disagree. If the honest answer is no, then nobody needed to read it, and no amount of punctuation surgery will fix that.

How to make AI writing sound human: put something at stake

The fix is not a writing technique. It is a decision about what you are willing to put your name next to. Five moves, all of them things a model cannot do for you:

Name the tradeoff you accepted. A platform engineer who writes "we moved to managed infrastructure and gave up the tuning we used to do by hand, and here is the month that hurt" has written something no model could invent, because it happened to them.

Say who your advice is wrong for. A hiring manager who says "this interview loop is a bad fit if you are hiring under ten people a year" is more credible in one clause than in three paragraphs of qualification.

Bring a receipt only you have. A fractional CFO who has closed forty sets of books knows which month the process breaks and why. A materials researcher knows which supplier spec is optimistic. Specificity of that kind is not retrievable; it is earned.

Say the unpopular half out loud. If your real position is that most of your industry's favorite practice is theater, write that sentence, then defend it. The defense is the article.

Keep the part that did not work. The failed pilot, the rollback, the client who left. Models produce clean arcs. People have receipts with corrections on them.

None of this is voice in the cosmetic sense. AI imitates voice well and gets better at it every quarter. What it cannot do is decide which claim you are prepared to be wrong about in public. That is judgment, and judgment is the half of writing that still has to come from you.

Why this matters now

The share of machine-drafted professional content is no longer marginal. Originality.ai's rolling analysis of LinkedIn found that more than half of long-form posts were likely AI-generated, and its more recent reading of long-form specifically puts the figure dramatically higher — with platforms now experimenting with signals that mark posts as likely slop. The practical consequence is not that readers will catch you. It is that safe, balanced, unobjectionable writing has become the default output of the entire feed.

Which means the old defensive instinct is now the risky one. Hedging used to make you look measured. In a feed where competent-and-agreeable is free and infinite, hedging makes you look automated. The strongest proof of authorship you can offer is a position someone could hold against you, because that is the one thing nobody generates by accident.

We built Prowir on that premise: the model can carry the drafting, but the opinion has to be yours, and the work is getting it out of your head and onto the page in a form that still sounds like you. The beta is open if you want to try it.

Stop auditing your dashes. Go write the sentence you would have to defend.