Common AI Writing Tells (and How to Remove Them)

Here's a sentence a language model handed us during testing:

It's important to note that automation can sometimes be a factor that may, in certain cases, contribute to improved efficiency across teams. On one hand, it saves time; on the other hand, results can vary depending on implementation.

Here's the same point after a five-minute edit:

Automation saves time when you automate the boring, repeated work. It backfires when you automate a process nobody's fixed yet, because now the broken version just runs faster.

Same claim. One version hedges, pads, and refuses to land anywhere. The other one says something you could agree or disagree with. That gap is what AI Writing Tells are, and this article walks through the specific, nameable patterns that make writing sound machine-made, plus the edit that closes the gap between draft one and something a person would actually publish.

Start With a Weak Draft

Every AI writing tell starts life as the safest possible choice. A model predicts the next likely word, sentence by sentence, and the likeliest continuation is almost always the most generic one, so a raw draft hedges where a writer would claim and lists where a writer would argue. None of that is a bug. It's what "average" looks like when you generate it at scale.

We build an AI writing product, and our team edits AI-generated drafts every day as part of that work. We run those drafts through SEO and citation scoring before a human ever reads them, which is where most of the observations in this piece come from, not from a formal study but from repeated hands-on editing. We see the pattern constantly, and not just in customer drafts. Our own generation pipeline scores every piece for SEO and for how well it might get cited by AI answer engines, then rewrites it up to twice before a human sees it. For a stretch of months, that second rewrite pass carried none of our brand voice rules. It was optimizing for search and citation signals. Quietly, it sanded off everything that made the writing sound like us. Nobody on the team noticed from inside the pipeline. A customer noticed from outside, reading the blog, and told us.

A weak draft isn't a moral failure of the tool. It's the default state before a specific person with a specific point of view has pushed back on it. The first draft above wasn't wrong about automation. It just hadn't decided anything yet. Naming the patterns that show up in that undecided state is the first step. Editing them out is the second, and it's a lot faster than most people expect once the patterns have names.

The Tells That Keep Showing Up

The tells that get called out most often, by readers and by our own editors, are a colon followed by exactly three parallel items, the "it is not X, it is Y" contrast sentence, and verbless fragments dropped in for punch. Readers flagged all three in our own published work, and they were right every time.

The colon-plus-three construction shows up as "the fix is simple: audit, edit, publish" when a full sentence would carry the idea better and without the tic. The contrast scaffold shows up as "this isn't about speed, it's about consistency," a shape that sounds decisive but usually just restates one idea twice. Verbless fragments show up as "One tool. One workflow. Done." which reads as confidence the first time and as a nervous habit the third.

It helps to sort these by what they're actually doing, because they don't all fail the same way. Hedging tells soften a claim until it means nothing, the "may in certain cases contribute to" family. Structural tells impose a shape the content doesn't need, the colon-plus-three list, the symmetrical section-by-section template, the reflexive closing summary. Rhythm tells are about pacing rather than wording, most obviously the uniform-sentence problem: three or four short declaratives in a row, none longer than the last, which reads as confident on the surface and mechanical underneath. Provenance tells sit outside the sentence entirely, things like a watermark or a model signature a platform checks against the file, regardless of how the prose reads.

Older, more familiar tells sit alongside the newer three. Filler phrases like "it's worth noting" clear a throat before saying nothing. Transitions like "moreover" and "additionally" imply a connection without making one. Inflated words like utilize and leverage stand in for use. False balance lists two sides and declines to pick one.

One of the most consistent tells we've hit is model-specific and stubborn in a way a prompt can't fix. We asked our model, repeatedly and in different phrasings, to avoid em dashes in generated copy. It kept using them anyway. We stopped asking nicely. We now strip them in code, deterministically, after every generation, before a human ever sees the draft. Style rules that actually matter need enforcement, not a reminder in the prompt.

We ran a script across our own 80 published articles checking for these exact patterns. The method was simple: count colon-plus-three constructions, contrast scaffolds, and a fixed list of hedging and filler phrases per thousand words, then flag any piece that cleared a threshold on more than one category. The threshold was a judgment call, not a published standard, and 80 articles is our own archive rather than a representative sample of AI writing generally. Within that scope, every single article failed at least one check. The worst piece hit roughly 13 colon-plus-three-item constructions per thousand words, one every few sentences. That's not a fluke in a rushed draft. It's what happens when nobody's specifically hunting for the pattern, because a model will reach for it by default and a tired editor will wave it through.

How to Edit the Tells Out

Removing these tells is a single, focused editing pass, not a rebuilt workflow. Run it separately from your fact-check and structure passes. It needs a different kind of attention, hunting for shape and rhythm rather than accuracy. In our own editing work, this pass has been enough to take a flagged draft down to zero failed checks on the script above, though that's a result from one workflow, not a claim about editing generally.

A quick pass looks like this:

1. Search for a colon followed by a short list. Ask if a plain sentence carries the same point with less scaffolding.
2. Check every "it is not X, it is Y" sentence. If it's just repeating an idea for effect, cut half of it.
3. Flag verbless fragments. Decide if each one earns its place as a deliberate beat, or if it's a nervous habit dressed up as punch.
4. Delete filler phrases like "it's important to note." The sentence almost always stands on its own without it.
5. Cut transitions that only announce "here's another sentence." Keep the ones marking a genuine turn.
6. Swap inflated words for plain ones. Utilize becomes use, facilitate becomes help.
7. Read the page shape. If every section runs the same length with the same rhythm, break it on purpose. Let one point sprawl and another sit in a single line.
8. Where the draft hedges between two options without landing anywhere, make the call the evidence actually supports.

That checklist catches style. It won't catch provenance, and the two aren't the same problem. A surface tell is something a reader notices in the sentence itself, a hedge, a fragment, a template shape. A provenance check is something a platform runs against the file itself, a watermark or a model signature, and it works whether or not the prose reads smoothly. The checks also differ by platform. Some answer engines weigh citation and sourcing signals more heavily than style, others lean on watermark detection like SynthID, and a piece can pass one kind of check while still tripping another. You can clean out every tell on the list above and a provenance check can still flag the piece, because it's checking a different layer entirely.

None of this needs new tooling. A checklist you run every time beats a clever prompt you run once, because prompts drift and checklists don't. Semrush's 2024 survey on AI content and SEO found that 73% of respondents already combine AI drafting with human editing, and only 9% reported worse SEO results from AI-assisted content. That lines up with what we'd expect: the editing step is where the difference gets made, not the generation step. OpenAI's own research on ChatGPT usage backs that framing too. Most writing sessions are requests to revise something already on the page rather than requests to generate from nothing. Writing, in practice, is mostly editing.

Detection is also moving past a reader's eye for style, and Google's own systems are the clearest evidence of that shift. Google's SynthID system has already watermarked more than 10 billion pieces of content and can scan text, audio, video, or images for that signature. The company says it's rolling out further internal signals in 2026 to flag AI-generated content automatically. Style tells still matter for the reader in front of you, but the ground underneath them is shifting toward provenance and platform-side checks running alongside the eyeball test.

For the full run-it-every-time version of this pass, our editing checklist for AI drafts covers voice and cutting in more detail than fits here. The broader case for why a human pass matters at all is in our piece on quality and editing. If you're optimizing for how AI answer engines cite your work rather than just how readers respond to it, generative engine optimization is the more specific angle. And if you're evaluating writing you didn't produce yourself, the same patterns read from the outside in our guide to spotting AI-written content.

Naming these AI Writing Tells doesn't fix a draft by itself. The fix is still the edit, done deliberately, on a schedule, until the patterns stop showing up without you having to look for them.