10 Tells That Give Away AI Writing

Ten patterns that make text read like it came out of a machine. Here's each one, plus a ready-made set of rules so you're not hunting for them by hand.

The False Contrast

"Not a tool, a revolution. Not speed, quality." It looks like a sharp, deliberate line — but it's rigged. It feeds you the wrong answer first, then the right one. Just write the right one.

Corporate-Speak

Plain verbs get swapped for noun phrases: "pay" turns into "process the payment," "decide" turns into "make a determination." That's the language of contracts and internal memos. Leave it there.

George Orwell called this out eighty years ago in "Politics and the English Language": the verb gets pushed out by a noun built from it, a chain of prepositional phrases piles on behind it, and the sentence stops moving.

Clichés

"It cannot be overstated." "It's worth noting." "In today's fast-paced world." This is term-paper padding — the stuff people write to hit a word count and put the grader to sleep.

Abstractions

Transformation. Efficiency. Scaling. Synergy. Simple test: swap in a number or an actual action. If it doesn't fit, the word is empty.

Transitions to Nowhere

"There's one important thing to note here." "But that's not all." "Which brings us to." A paragraph of filler always follows. A paragraph should open with content — it doesn't need a bridge to the one before it.

Worn-Out Closers

"The choice is yours." "Only time will tell." "The future looks bright." These went stale a decade ago and land on exactly nobody now.

Hedges

"May." "In some cases." "Typically." "Potentially." Readers catch this instantly: no actual position taken. Either the claim is true — say it straight. Or it has a condition — name the condition.

Clickbait Headlines

"This one mistake will destroy your business." "One prompt will change everything." Works on people who just found the internet yesterday, or anyone still clicking on full-screen banner ads.

Padding

"Technology is evolving at a breakneck pace, which makes it hard to say anything for certain." Three lines, zero content. Would you finish reading that yourself?

The Flat Tone

A wall of text with no opinion in it, every sentence roughly the same length and the same temperature. The model averages by default, because averaging is its whole job.

Word Lists Go Stale

People edit their drafts knowing these lists exist, so any list of trigger words stops working within months. Lean on the shape of the text, not a checklist of forbidden words. A banned-word dictionary rots faster than you can finish compiling it.

A Ready-Made Ruleset

Cleaning this up by hand every time isn't necessary. There are open skills for AI agents that audit and fix text against a list of patterns. The most established one built for English sits here:

Repo: https://github.com/conorbronsdon/avoid-ai-writing

One warning before you wire it in — worth reading, not skipping. I haven't independently verified how well it performs. The "42 patterns" and the structured four-section audit are claims from the project's own README; nobody outside the project confirmed those numbers. The skill takes your whole draft and edits it against its own rulebook, so open the pattern list on GitHub and actually read it before you let it touch your writing.

The star count and last-commit date matter too — 1.5k stars, 157 forks, as of this write-up in August 2026. That's a snapshot, not a guarantee. A repo like this can go quiet tomorrow, and its pattern list ages right along with it. Same rule applies to any tool like this you find later: read the source before you trust the output.

Unlike the Russian-language side of this problem, English isn't short on mechanical linters either. Vale, write-good, alex, and proselint all ship with mature, actively maintained style rules — Google's and Microsoft's style guides are even available as ready-made Vale packages. They weren't built to catch "AI tells" specifically; they catch passive voice, wordiness, and biased phrasing. Still worth running alongside a purpose-built skill like the one above.

What Still Works From the Old Guard

Hemingway App has been around since 2013 and it's still standing. It flags passive voice, adverbs, and sentences too complex to parse in one pass, and hands you a grade-level readability score. The web version is free, no login required. The $19.99 desktop app adds offline editing and export; a paid subscription on top of that bolts on AI rewriting.

Same caveat applies here as with any readability score: it measures clutter, not quality. A perfect Hemingway grade and an empty paragraph get along just fine.

Workflow

• The ruleset skill lives inside the project — otherwise you're repeating the same request every single session
• After a draft generates, it runs through the pattern list automatically
• Whatever gets flagged gets fixed; everything else stays untouched
• Final read-through with your own eyes — a tell is a skew across the whole piece, and no machine catches that
• Revisit the pattern list every few months. It goes stale, same as the word lists.