18 Signs a Piece of Text Was Written by AI

You don't need a detector tool to spot AI-generated text. After reading thousands of drafts, editors develop a feel for it. The prose has a certain texture: polished on the surface, hollow underneath. Sentences that sound almost right but say nothing you couldn't have said better with fewer words.

Here are 18 specific patterns that show up again and again in AI output. Some are mechanical. Others are tonal. All of them are fixable.

The structural giveaways

These are the signs you notice before you even read closely. They're in the shape of the text.

1. The em-dash everywhere

AI models love the em-dash. Not one or two per article, but a dozen or more, often in spots where a comma or a period would serve the reader better. A single em-dash used well has force. A paragraph with four of them reads like the writer couldn't commit to any sentence structure at all. We've written a full breakdown of this specific tic in why AI loves the em-dash and how to spot it.

2. The rule of three, always

"It's fast, flexible, and powerful." "Clear, concise, and compelling." If you see triplets everywhere in a piece of text, that's an AI tell. Human writers occasionally reach for three things when listing genuinely calls for three things. AI defaults to three because it learned from marketing copy that triads sound authoritative. The rule of three as a reflexive habit (not a stylistic choice) is one of the more reliable AI writing patterns to watch for.

3. Every section the same length

Read a suspicious piece and count the paragraphs per section. If each H2 block has almost exactly the same number of sentences, you're probably looking at AI output. Human writing is uneven. A point that needs two sentences gets two. One that needs eight gets eight. AI models tend to pad short points and compress long ones until everything reaches a tidy equilibrium.

4. An introduction that restates the title

"In this article, we'll explore the signs that a piece of text was written by AI." If the first sentence could be auto-generated from the title (and it often can), that's a sign. Human writers have a reason to open the way they do. They're trying to earn your attention. AI is just warming up.

5. A conclusion that summarizes everything you just read

"In conclusion, by understanding these key patterns, you'll be better equipped to..." Humans don't write conclusions like this. They end with a thought, a tension, or a real question. Something with weight. AI circles back and recaps because it learned this structure from essays that required one. It's the written equivalent of "in summary, here is my summary."

The vocabulary tells

A set of specific words appear in AI writing at rates no human writer would hit. Spotting them is fast once you know what to look for.

6. The flagged words list

There's a cluster of words that AI models reach for constantly: delve, tapestry, testament, seamless, robust, realm, vibrant, pivotal, foster, garner, leverage (the verb), navigate (used metaphorically). If you see three or more of these in a single piece, you almost certainly have an AI draft. The words that instantly signal AI-generated text go deeper on this with frequency data.

7. "Serves as" and "stands as"

"This guide serves as a comprehensive resource." Just say it is one. "Serves as" adds nothing except a bureaucratic distance between the subject and its predicate. It shows up constantly in AI because it sounds formal without requiring the model to actually commit to a claim.

8. Vague attribution

"Studies show that..." Which studies? "Experts agree..." Which experts? "Research suggests..." From where? AI models hedge with vague attribution constantly because they can't reliably cite specific sources. A human writer who cares about credibility names the study, names the researcher, or at minimum admits they're speaking from experience rather than data.

9. "In today's fast-paced world" and cousins

Openers like "In today's world," "In our increasingly digital age," and "As technology continues to evolve" are AI tells. These phrases exist because AI learned to begin articles with context-setting statements. They say nothing. A real writer who starts with context is specific: "Since GPT-4 launched in 2023, the volume of AI-assisted copy online has increased by a measurable amount." Not just "technology is changing things."

The sentence-level patterns

10. Robotic uniformity in sentence length

Read a paragraph aloud. If every sentence takes about the same amount of time to say, the prose is likely AI output. Human writers vary rhythm naturally. Some sentences are four words. Some unfold over two clauses because the idea actually takes that long to develop. AI models produce metronomic text. Not because it's more correct, but because that's the safest path through training.

Before (AI draft): "Content marketing is an important strategy for businesses. It helps companies build brand awareness and attract new customers. Many organizations use it to drive long-term growth."

After (human rewrite): "Content marketing works. Slowly, then all at once. Companies that stick with it for a year or two tend to see compounding results that paid ads can't replicate. Those that quit after three months usually don't."

The rewrite has a point of view. It makes a claim a reader might disagree with. The original has none of that.

11. "-ing" tail clauses that explain the obvious

"She revised the document, ensuring clarity for the reader." The second clause is dead weight. Of course revising a document aims for clarity. AI writers add these trailing participial phrases to make sentences feel more sophisticated. They don't. They dilute. Cut them.

12. Negative parallelism

"It's not just about writing well. It's about writing with intention." "This isn't just a guide. It's a toolkit." Human writers occasionally set up contrast this way. AI does it constantly. The pattern "not just X, but Y" or "not only X but also Y" appears at disproportionate rates in AI output. It's a taught rhetorical move that the models apply reflexively rather than selectively.

13. Inflated symbolism

"The humble pencil is a powerful metaphor for creativity's relationship with impermanence." AI models, trained on content that signals depth, sometimes reach for symbolic meaning where none is warranted. A pencil is a pencil until the piece actually needs it to mean something. This kind of inflation is especially common in introductions and conclusions.

The argument-level tells

These are harder to articulate but important. They have to do with whether the text actually thinks.

14. Claims that don't commit

"This approach may offer several potential benefits in certain situations." Every hedge word is doing nothing except making the claim unkillable. Real writing takes a stance: "This works for short-form but falls apart in anything over 800 words." AI models hedge because extreme caution reduces the chance of a factual error. The cost is prose that sounds like a legal disclaimer.

15. Balance where the writer clearly has an opinion

"Some people find AI writing useful. Others prefer a more human touch. Both perspectives have merit." This kind of false balance shows up when an AI model is trying not to offend anyone. But the piece is supposed to have a point. Refusing to take a position isn't balance. It's a dodge.

16. Comprehensive coverage without depth

AI text tends to cover everything at the same depth: a paragraph per subtopic, no more, no less. Human writers spend more space on the parts that are harder, more interesting, or more consequential. If a piece feels like it was assembled from a listicle outline without anyone deciding what actually mattered, that's a structural sign of AI authorship.

17. No failure, no friction

AI-generated content rarely mentions things that didn't work. Human experience writing includes setbacks, qualifications, things the writer tried before finding the approach they're describing. When a how-to article reads as if every step was obvious and everything went smoothly, it usually means nobody actually did the thing.

18. The generic upbeat sign-off

"Start implementing these strategies today and watch your results soar." No. A real editor ends on something concrete: the next step, an honest caveat, a question the reader should ask themselves. "Sign-off inspiration" is a content-mill invention, and AI learned it.

What to do with these signs

Finding these patterns in a draft isn't a judgment. It's a starting point. Most of them are fixable in revision. The free humanizer prompt at /humanizer-prompt is built around these same signals: it gives the model specific instructions to vary rhythm, cut hedge words, drop the flagged vocabulary, and take actual positions.

The deeper fix is editing with intention. Read the draft as if you have to defend every sentence to a skeptical reader. Why this word? Why this length? What is this sentence actually claiming? Those questions cut through AI texture faster than any automated tool.

FAQ

How reliable are AI detector tools compared to reading for these signs?

Detector tools have real false-positive rates, and they change as models improve. Reading for patterns is slower but more durable. A piece can fool a detector while still reading, to a careful human, exactly like AI output. For high-stakes copy, treat the manual check as the primary review.

Can human writers accidentally use these patterns?

Yes. That's part of why they're worth knowing. A writer who defaults to rule-of-three structure, vague attribution, and hedged claims is writing like AI even if they never touched a generator. The patterns aren't inherently AI. They're signs of writing that doesn't commit. AI just produces them at industrial scale.

Does fixing these patterns make the text harder to detect?

That's a secondary concern. The primary reason to fix them is that the writing becomes better: more specific, more honest, more readable. Whether or not a detector changes its score is beside the point.

What's the single fastest sign to check for?

The em-dash count. Open the document, search for "—", and count. More than three in a standard article is a flag. More than six is almost always AI. It takes five seconds to check and has a low false-positive rate.

Are some AI models better than others at avoiding these patterns?

Some. But all of them have tendencies, and those tendencies shift with each model update. The patterns described here are structural and tonal, not tied to a specific model's vocabulary list. They'll remain useful as the models evolve.