The 7 AI writing tells editors catch first

Seven patterns give a machine-drafted page away to an experienced editor: em-dash bombardment, stock phrases, autopilot triads, uniform sentence length, an absent author, symmetrical hedging, and a conclusion that restates the introduction. Five of them are countable. Two need a reader.

This is not a sneer at people who write with AI. It is the checklist I run against my own copy, and my drafts start dirtier than I would like to admit. On September 21, 2026 I ran the first five rules across every post on this blog, 21 posts and about 34,000 words, and eight of them were over my own em-dash budget. The worst sat at 9.8 per thousand words against a budget of four. Fixing all eight took a morning, which is the actual point of having a countable rule.

The first five are mechanical enough that we built them into a free AI writing checker that flags each one with its exact position. The last two you have to catch by reading.

1. The em-dash bombardment

Language models reach for the em-dash as a universal connector that stands in for a comma, a colon or a parenthesis in one keystroke. One or two per paragraph reads stylish. Eight per page reads generated, and editors now flinch at the pattern on sight. Our budget is four per 1,000 words, and most flagged dashes are better as a period anyway. The short sentence that results is usually the stronger one.

2. Stock phrases nobody says out loud

Delve into. In today's fast-paced world. It's worth noting. Moreover starting a sentence that did not need a connector at all. Each phrase is grammatical, and that is the problem: they are the statistically safe choice, which is exactly what a model optimizes for and exactly what a human voice does not sound like. Read the sentence aloud without the phrase. If it survives, the phrase was filler. If it dies, the sentence was.

This one is measurable at scale. A study of more than 15 million PubMed abstracts found that certain style words, "delves" among them, jumped in frequency right after LLMs arrived. The authors estimate at least 13.5% of 2024 biomedical abstracts were processed with an LLM.

3. Triads on autopilot

Fast, clean, and simple. Smart, driven, and kind. The rule of three is real rhetoric with real power, which is why models deploy it constantly and why the fifth triad on a page reads like a template stamping itself. One triad is a choice. Five is a habit. Cut the weakest item from most of them; a pair reads like a decision was made.

4. Sentences all the same length

This is the strongest single tell, and the least discussed. Models regress toward medium-length sentences of suspiciously similar shape. People do not write that way. They fragment. Then they pile clauses into something that runs longer than it should because the thought refused to end where the grammar wanted it to. Measured as variation in sentence length, human prose usually sits well above 35%; flat drafts sit below it. The fix is surgical: cut one sentence to four words, let another sprawl, and read the paragraph aloud.

5. Nobody home

Eight hundred words and not one I or my anywhere in them. No date, no mistake, no opinion held against interest. Voiceless prose is not proof of machine authorship, committees produced it for decades, but it fails the same way: the reader cannot locate a person who stakes anything on the claims. The cheapest fix in this whole list is one sentence only the author could write. What it cost. When it broke. What you would not do again.

6. Hedging without stakes

Here is where the mechanical checker stops and editorial judgment starts. Machine drafts hedge symmetrically: results may vary, it depends on your use case, there are tradeoffs to consider, with no commitment about which way the tradeoff usually falls. Human experts hedge asymmetrically, because experience has weight: this usually fails for B2B, although I have seen two exceptions. If a paragraph considers every possibility and recommends nothing, no rule will flag it, and every experienced reader will.

7. The conclusion that restates the introduction

Models close essays the way they were trained to, by summarizing what was already said, often beginning with In conclusion. A human who actually went somewhere ends somewhere new, with the implication, the next action, the cost of ignoring all of it. If your last paragraph could be deleted with no information lost, it already has been, by every reader who stopped at the previous one.

Why the countable five are worth automating

The split between the first five tells and the last two is not arbitrary, and it decides how you should spend your editing attention.

The countable ones are cheap to fix and expensive to notice. Nobody reading a draft thinks "that is 9.8 em-dashes per thousand words". They think the page feels slightly off and they cannot say why, which means the feedback never reaches you as feedback. A rule that runs in a second turns an invisible, slow-bleeding problem into a list with line numbers.

The uncountable two work the other way round. Hedging without stakes and a conclusion that restates the introduction are both obvious the moment a reader notices them, and neither can be flagged by a rule, because the same sentence is correct or empty depending on whether the writer knows something. No tool will ever tell you which.

So automate the five and read for the two. The common failure is the reverse: people run a detector, get a score, and treat prose quality as settled.

Are AI detectors reliable enough to judge a draft?

No. I refuse to run AI detectors on principle and on math: the same text scores wildly differently across services, and false accusations are a documented harm. A Stanford-led study found popular GPT detectors consistently misclassified essays by non-native English writers as AI-generated, while native samples passed. Detectors grade authorship. Editors grade prose. The seven tells measure something more honest: whether the prose earns trust. That is checkable and fixable regardless of who typed the draft.

Google draws the same line. Its guidance on AI-generated content rewards helpful content however it was produced, and its scaled content abuse policy targets mass-produced pages made to game rankings, with or without AI. The tells above are how that low-value prose usually shows up on the page.

When I rewrote all nineteen product descriptions for this catalog, this checklist was the gate, and the workflow that consistently passes it is packaged in Humanizer Pro. The rewrite story, including before-and-after copy, is in rewriting 19 product descriptions in 2 days. Paste your own draft into the checker first. The flags it raises are the same five I still catch in my own writing, which is the most honest endorsement a checklist can have.