Straight answer: AI writing gives itself away through a small, repeatable set of patterns: sentences built on “it’s not just X, it’s Y,” lists that always land on exactly three items, transition words like moreover and furthermore doing heavy lifting, paragraphs that are all roughly the same length, and endings that neatly summarise a point nobody asked to have summarised. Once you have spotted three or four of these in the same piece of text, you are not looking at a coincidence.
Why I notice this more than most people
I read a lot of drafts. Client blog posts, guest post pitches, LinkedIn updates people send me before they hit publish, cover letters from people who want to work with me. Some weeks it is thirty pieces of text, some weeks it is sixty. After a few years of that, the patterns stop feeling like patterns and start feeling like a fingerprint. You see the same shapes over and over, from people who have never met each other, using different tools, writing about completely different subjects.
That is the part that took me a while to accept. It is not that AI writing sounds bad. Some of it reads perfectly cleanly. It is that it sounds the same, sentence structure and all, regardless of who typed the prompt. A dentist in Leeds and a SaaS founder in Austin will produce blog intros with the identical rhythm if they both leaned on the same tool with the same default settings. Humans do not naturally converge like that. AI does.
The sentence-level tells
These are the ones you can catch inside a single sentence, no context needed.
- The false contrast. “It’s not just about productivity, it’s about peace of mind.” This construction shows up constantly because it sounds profound while saying almost nothing. Real writers usually just state the second half.
- The rule of three, everywhere. Three benefits, three reasons, three tips, three adjectives in a row (“fast, flexible, and affordable”). Threes read well, which is why AI models were trained on so much writing that uses them, which is why they now overuse them to the point of self-parody.
- The hedge that adds nothing. “It’s worth noting that,” “it’s important to remember,” “one could argue.” These phrases exist to sound careful. They usually precede a sentence that did not need the warning.
- The summary sentence that restates what you just read. “In short, these strategies can help you achieve your goals.” Nobody who wrote that paragraph needed to be told what it said thirty seconds earlier.
The structural tells
Zoom out from the sentence to the whole piece and different patterns appear.
- Paragraph symmetry. Every paragraph is three to four sentences. None run long, none run short. Real writing breathes, some points need one line, some need eight.
- Headers phrased as questions or commands. “Why Does This Matter?” “How to Get Started” repeated section after section, like a template that was never customised for the actual subject.
- The conclusion that adds a fresh, unearned rallying cry. A closing paragraph that suddenly gets motivational, telling the reader to “take the first step today” when nothing in the previous 1,500 words was about urgency.
- Even-handedness on everything. AI writing loves to present “both sides” even on things that do not have two sides. It hedges where a human with an opinion would just tell you the answer.
The word-level tells
Certain words show up disproportionately in AI-generated text compared with how often people say them out loud or type them in normal writing. Watch for: moreover, furthermore, additionally, boasts, testament, tapestry, landscape (as in “the marketing landscape”), realm, “in today’s fast-paced world,” “when it comes to,” “dive into,” “unpack this,” and “at the end of the day” used as a throat-clearing opener rather than an actual conclusion. None of these words are wrong on their own. The problem is density. One “moreover” in a 2,000-word article is fine. Four in 900 words is a signal.
A real example from my inbox
Last year I was helping a small business owner hire a virtual assistant to handle emails and light content. Thirty-one people applied. I could tell within the first two lines which cover letters had been generated, not written, by the applicant. The tell was not one big obvious thing. It was the accumulation.
One application opened with “I am thrilled at the opportunity to potentially contribute my skills to your esteemed organisation.” Nobody talks like that. Then it moved into a three-part list of “why I am a strong fit,” each point exactly two sentences, each sentence structured the same way: a claim, then a slightly grand restatement of the claim. It closed with a paragraph about being “excited to embark on this journey together.” Four separate patterns in 200 words. That candidate was rejected within about ninety seconds of opening the email, not because using AI to draft an application is a crime, but because it told me the person had not bothered to read what they sent, let alone edit it into something that sounded like them.
Compare that to another applicant, who wrote: “I’ve spent two years doing customer support for a small ecommerce brand, mostly refund requests and shipping complaints. I got fast at de-escalating annoyed customers because I had to be, the founder had zero patience for angry emails sitting unanswered.” That is specific. That has a real detail (refund requests, an actual frustration, a real reason). You cannot fake that texture without having lived it, and it is the exact kind of detail I go into in more depth when I talk about what works in customer service remote job applications, where the hidden pattern is almost always specificity, not polish.
Why “sounding more human” often backfires
Here is the part most people writing about this topic will not say out loud. Once writers learn these tells, a lot of them start deliberately faking human imperfection. They throw in a sentence fragment. A rhetorical question. A slightly odd sentence length on purpose. And it usually reads worse than the AI original, because forced imperfection has its own pattern too, and readers pick up on strain even when they cannot name it.
The uncomfortable truth is that plenty of skilled human writers naturally use rule-of-three lists, short punchy paragraphs, and the occasional summary sentence, because those are just useful tools that predate AI by decades. Copywriters have used “not just X, but Y” since long before any language model existed. If you go hunting for these patterns to accuse a real person of using AI, you will catch honest writers as often as you catch anyone gaming the system. I have seen AI detection tools flag entirely human-written cover letters as “likely AI generated” with high confidence, and I have seen heavily AI-assisted articles sail through undetected because someone spent twenty minutes rewriting the transitions. The tools guess. They do not know.
What matters, especially if you run a business and use AI in your content the way most of us do now, is not whether a tool touched the draft. It is whether the final version says something specific that only you could say. I write plenty of first drafts with AI. Then I go through and rip out every generic claim and replace it with a number, a name, a date, something from my own week. That is the whole job now. If you want a proper structured way to get good at that instead of fighting AI or pretending you never use it, I go through the exact process in how to become fluent in AI in 90 days.
The seven-step check I run before anything goes out
Whether I am publishing a blog post, reviewing a client’s draft, or sending a proposal, this is the pass I do:
- 1. Read it out loud. Anything that sounds like a corporate press release when spoken gets rewritten on the spot.
- 2. Count the threes. If more than one list or sentence uses exactly three items, cut one to two or expand one to four.
- 3. Search for “not just” and “it’s not about.” Delete the sentence or flip it into a direct statement.
- 4. Find and remove moreover, furthermore, additionally. Ninety percent of the time the sentence works fine without them.
- 5. Check paragraph lengths. If five paragraphs in a row are the same length, break the rhythm deliberately.
- 6. Add one detail nobody could have guessed. A real number, a real client story, a real date. This single step does more to make writing sound human than every other step combined.
- 7. Cut the closing summary paragraph. End on the last real point instead of a paragraph that repeats the article back to the reader.
This same discipline applies whether you are writing a blog post or crafting a headline that has to earn a click in half a second. If you want more on that specific skill, the structure in writing more effective headlines covers exactly the kind of specificity that separates a headline that sounds like a template from one that pulls a reader in.
Where this shows up beyond blog posts
It is not just blogs and cover letters. I see it in job listings for home-based writing and admin work, where the ad itself was clearly AI-generated (three bullet points of “benefits,” a closing line about “joining our dynamic team”), which is worth knowing before you apply, because a listing full of these tells is often a sign nobody at the company has reviewed it either. I wrote about spotting the legitimate opportunities from the templated ones in typewriting jobs from home, and the same instinct applies there as it does in cover letters: specific detail beats polish every time.
It also shows up now in automated content operations, where businesses run entire content pipelines through an AI agent with no human check at the end. I am not against building those, I help clients set them up, and if you want to see how far you can take this without writing a line of code, where to build your own AI agent without coding is a good place to start. The problem is not the agent. The problem is publishing whatever it produces without the seven-step pass above.
The pattern behind the patterns
If there is one thing I would want you to take from all of this, it is that every single tell on this page comes from the same root cause: AI models are trained to produce the statistically most likely next sentence, which means they gravitate toward whatever pattern shows up most often across billions of words. That makes AI writing competent and forgettable at the same time. It is never wrong. It is rarely interesting. Human writing, at its best, includes the odd detail nobody else would have thought to include, because it happened to you specifically. That is the one thing no amount of prompting can manufacture, and it is the one thing worth protecting in your own writing, whether AI touched the first draft or not.
Frequently asked questions
Is using an em dash a sign that something was written by AI?
On its own, no. Plenty of professional human writers use dashes constantly and always have. It only becomes a useful signal when it shows up alongside three or four other patterns, like the rule of three, moreover-style transitions, and a neatly summarised closing paragraph, in the same piece.
Can AI detection tools reliably prove something was written by AI?
No. These tools work on probability, not proof, and they regularly flag human-written text as AI-generated while missing heavily AI-assisted text that has been lightly edited. Treat a detection score as one data point, never as evidence on its own.
Should I stop using AI to write if these patterns are so easy to spot?
No, the patterns are easy to spot in unedited AI output, not in AI-assisted output that a person has reworked. Use AI for the first draft if it helps, then run it through a proper editing pass that adds real detail and cuts the generic transitions and summary lines. That is the difference between text that reads as templated and text that reads as yours.
Why do AI tools keep producing lists and points in groups of three?
Because the training data they learned from is full of rule-of-three writing, a technique that has been used in speeches, advertising, and journalism for decades because it is satisfying to read. AI models picked up the pattern and now overuse it far past the point where it still feels natural, which is exactly what makes it noticeable when you know to look for it.
