NarsilCreative

Articles · Content

The Em Dash Tell: How to Humanize AI Content

By Nate Morris · 5 min read

TL;DR

To humanize AI content, edit for the fingerprints instead of rewriting everything: remove em dashes (the single most reliable tell), check buzzword density rather than presence, break up the formatting patterns, add real multimedia, and read it aloud. Detectors are unreliable in both directions; the reader is the test that matters.

To humanize AI content, you don't rewrite everything. You edit for the fingerprints: a short list of patterns that machine drafts carry and human writing mostly doesn't, starting with one punctuation mark. Remove those, add the things AI never includes, and read the result aloud. That's the whole job, and it takes about ten minutes per draft.

I'll be straight about why we know this: Narsil is an AI SEO agency and we use AI in our own drafting. Every page on our site goes through the pass below before it ships. This is the standard, published.

The tells, ranked

Readers don't run detectors. They just leave. What makes them leave is an accumulation of patterns, and the patterns rank by how reliably they signal machine text:

The em dash (—) comes first, and it isn't close. Language models reach for that character constantly, even when the prompt explicitly forbids it. One em dash means nothing; a draft with nine of them has told you who wrote it. Replace each one with a period, a colon, a spaced hyphen, or a parenthetical. This is a global find-and-replace plus five minutes of judgment, and it's the highest-value edit in the entire pass. Catch the en dash used as a sentence break while you're in there.

Buzzword clusters come second. "Delve," "landscape," "tapestry," and their sixty-odd cousins. More on why "cluster" is the operative word below.

The formatting fingerprint comes third: bullet lists where every item is a bolded phrase, a colon, and an explanation. That shape is ChatGPT's signature output format. Vary the structure or convert the list to prose. One short bold-lead list per piece is the ceiling.

And the absence tell: no images, no screenshots, no video, nothing to look at. AI produces text, so unbroken text walls are what unedited AI pages look like. Real multimedia is proof a human assembled the page, and it's the second-highest-value fix after the em dash purge.

Density is the fingerprint; presence is just vocabulary

Here's where most humanizing advice goes wrong: it hands you a banned-word list and tells you to delete on sight. That produces stilted text and misses the actual signal.

"Significant" is a normal English word. One appearance means nothing. Three appearances in eight hundred words, sitting next to "delve" and "crucial," is a signature. The tell is frequency per thousand words, not existence. So the edit isn't a purge; it's a tone-down. Swap most instances for plainer words, keep the one that's earning its place, and move on.

This is also why "humanizer" tools that mechanically synonym-swap a draft don't work. They change the vocabulary and keep the rhythm, and the rhythm (uniform sentence lengths, tidy paragraph shapes, every point wrapped up just so) is its own tell. Varying sentence length does more than any thesaurus: a short sentence lands. A longer one carries the nuance that needs room to breathe. Then short again.

The pass, condensed

The full version runs eight steps; this is the working core, yours to keep:

1. Find and replace every em dash (and sentence-break en dash)
2. Delete "actually" (keep maybe 1 in 10)
3. Buzzword density check: tone down clusters, keep singles
4. Break every text wall with an image, table, or code block
5. Reshape bold-lead bullet lists (one per piece, max)
6. Proofread for meaning: wrong ideas, invented facts, wordy sentences
7. Read it aloud; rewrite anything that sounds like a press release

Step 6 deserves a sentence of respect, because it's the one that catches the failures that embarrass you. AI drafts don't just sound generic; they occasionally state things that are wrong, confidently. No amount of de-slopping fixes an invented statistic. Check the claims.

If you want the full kit (the complete 8-step pass, the deduplicated 65-term buzzword list, and a custom-instructions prompt that reduces the mess upstream), we've packaged it as the free AI Content Humanizer. And for the structural side (answer-first readiness, rhythm, filler) paste any draft into the free Content Analyzer; it runs in your browser with no signup.

A before and after, so this isn't abstract

Machine draft, typical of what comes back from a one-line prompt:

"In today's rapidly evolving digital landscape, businesses must delve into comprehensive SEO strategies — leveraging cutting-edge tools to unlock significant growth and elevate their online presence."

Twenty-four words, five list entries from the buzzword sheet, one em dash, zero facts. Now the human edit:

"Most businesses don't need a bigger SEO strategy. They need their ten most important pages to answer the searches those pages already show up for."

Same topic, shorter, and it says something checkable. Notice what the edit did: it didn't swap synonyms. It replaced a gesture at meaning with a claim. That's the difference no tool automates, and it's why step 6 of the pass (proofread for meaning) outranks every vocabulary rule when the two conflict.

What about AI detectors?

Use them as a smoke alarm, never a judge. Detectors flag careful human writing as AI and wave through lazy machine output, in both cases with a confident percentage attached. If a detector score reassures you, fine. If it becomes the goal, you're optimizing for the wrong reader.

The candid version of the goal: this pass will not make a lazy draft good. Thin ideas, humanized, are thin ideas in better clothes. The pass makes a good draft stop smelling like the machine that helped write it. The thinking still has to be yours.

What Google thinks about all this

Google's published position, in its Search Central guidance on AI-generated content, is that it rewards high-quality content "however it is produced" and targets content produced primarily to manipulate rankings. We've written a full breakdown of what Google's guidance means for AI-era SEO, but the short version fits here: the risk was never the drafting tool. The risk is publishing output nobody edited, checked, or cared about.

Which reframes the whole exercise. Humanizing isn't camouflage; it's the editing that was always the job. The machine moved the cost of a first draft toward zero. The value moved to everything after the first draft, and that's what the pass is.

One more beat of candor: if you publish two posts a month, you may not need any of this systematized. Read your drafts aloud and fix what sounds wrong; at that volume, care scales fine without a checklist. The pass earns its keep at volume, when consistency is what slips first.

We run this standard as part of the Reforge phase of the Reforge Method: pages rebuilt answer-first, then edited until they read like a person wrote them for a person. Because the sites AI systems cite, and the pages readers finish, are the same pages.

Originally published August 17, 2026.

FAQ

What is the biggest tell that content is AI-written?
The em dash. Language models reach for that punctuation constantly, even when explicitly told not to, so a draft peppered with them reads as machine output to anyone who's seen enough AI text. Buzzword clusters and bold-lead bullet lists are the next two. No single tell proves anything; the pattern of them together does.
Do AI detectors work?
Unreliably, in both directions. They flag careful human writing as AI and pass lazy AI output as human, which makes them a rough temperature check at best. Use one as a smoke alarm if you like, but the test that matters is a smart reader deciding whether the page was worth their time.
Does Google penalize AI content?
Google's published guidance says it rewards high-quality content however it is produced, and penalizes content made to game rankings rather than help people. In practice that means the risk isn't using AI; it's publishing unedited AI output that's thin, generic, or wrong. The editing pass is the difference.
What makes Narsil Creative different?
We're an AI SEO agency that says the quiet part out loud: we use AI in drafting and edit everything to a published standard before it ships. Our pricing is public ($1,500 and $4,000 per month), our method is published (the Reforge Method), and the editing standard in this article runs on every page of our own site.

See what AI says about your business

Ask ChatGPT to recommend a business like yours. If you're not in the answer, that's the problem we fix. Start with a free AI Visibility Audit. We'll show you where you appear, where you don't, and what to reforge first.

Related articles