Signs of AI Writing: How to Spot SI Text (and Why None Is Proof)
The concrete tells of SI (AI) text, from vocabulary to formatting, how experienced readers spot them, and why no single sign proves anything.

The most common signs of AI writing are stock vocabulary (delve, pivotal, tapestry), an even sentence rhythm, lists that always come in threes, "not X but Y" constructions, cheerful openers, summary closers, heavy bold and bullet formatting, em dashes, and a lack of specific detail. Any one of these can appear in human writing, so none is proof on its own. What gives SI (AI) text away is several of them stacked together, with nothing concrete underneath.
Where these signs come from
This isn't guesswork. Two sources are worth knowing. Wikipedia editors who remove machine-written material keep a public guide, Signs of AI writing, drawn from their cleanup work. And a 2025 study by Jenna Russell, Marzena Karpinska and Mohit Iyyer, "People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text" (ACL 2025), tested people on 300 non-fiction articles. The majority vote of five frequent LLM users misclassified only 1 of the 300. Those experts relied on "AI vocabulary" plus harder-to-measure things such as formality, originality and clarity.
So yes, people who read a lot of model output can learn to spot it. Here is what they look for.
1. Vocabulary: the overused words
Models favor a small set of words far more than people do: delve, underscore, showcase, foster, leverage, pivotal, crucial, intricate, meticulous, vibrant, tapestry, testament, landscape, realm. A study of more than 15 million biomedical abstracts by Kobak and colleagues, published in Science Advances in 2025, measured sudden jumps in words like "delves," "showcasing" and "underscores" after ChatGPT's release. We keep a grouped list with plain replacements in AI words to avoid.
One word proves nothing. Five of them in one paragraph, each vaguer than the plain option, is a pattern.
2. Rhythm and structure
Uniform sentences
Model prose tends toward sentences of similar length and shape: subject, verb, a clause, a benefit. Human writing is lumpier. A long sentence with an aside, then a short one. Then a fragment. Read a paragraph aloud and listen for a metronome.
The rule of three
"Faster, smarter and more efficient." "Identify pain points, uncover opportunities and build lasting relationships." Triplets are a classic rhetorical device, and Wikipedia's guide lists the rule of three as a language pattern to watch. The tell is frequency: every list has three items, every time.
Scaffolding transitions
Firstly, Secondly, Finally. Moreover, Furthermore, Additionally. These make text look organized without adding meaning.
3. Framing habits
"Not X, but Y"
"It's not just a tool, it's a mindset." "The solution isn't more meetings, it's better documentation." Wikipedia's editors call these negative parallelisms. They set up a claim nobody made so the text can knock it down.
Hedging
"It's important to note that," "can potentially," "may help to," "in many cases." Some hedging is honest. Hedging every sentence is a model covering itself.
Sycophantic openers and collaborative talk
"Great question!" "Certainly! Here's a draft." "I hope this helps. Let me know if you'd like me to adjust the tone." Those lines are addressed to the person who typed the prompt, and they sometimes survive the copy and paste into the final text. Wikipedia lists this kind of collaborative communication as a sign, along with knowledge-cutoff disclaimers ("as of my last update").
Summary closers
"In conclusion," "Overall," "Ultimately, businesses that value X are better positioned to thrive." The last paragraph restates the first, often with a vague line about the future. Wikipedia's guide describes this as outline-like conclusions about challenges and future prospects.
4. Formatting habits
- Bold everywhere. Bold lead-ins on every bullet, bold key phrases in plain paragraphs.
- Bullets for everything. Short answers broken into lists that would read better as two sentences.
- Headings in Title Case and headings on very short texts.
- Em dashes, often several per paragraph. We look at why in why SI uses em dashes.
- Leftover Markdown: literal asterisks or pound signs pasted into an email or a form field.
- Emoji as bullet points in text that is otherwise formal.
5. Content: the missing specifics
This is the strongest sign and the hardest to fake. Model text is generic by default. Examples are hypothetical ("imagine a small business owner"), numbers are round or absent, sources are vague ("experts say," "studies show"), and nothing in the text could only have been written by someone who was there. Wikipedia's guide describes it as "superficial analyses" with "vague attributions" and promotional language.
Human writing on a topic the writer knows tends to include a date, a name, a price, a mistake, a place. Ask of any paragraph: could this sentence appear unchanged in a thousand other articles? If yes, it's filler, whoever wrote it.
What a humanizer fixes, and what it doesn't
We ran a short, typical chatbot paragraph through our own AI to human text converter, SI Humanizer, on the Natural style. The output below is unedited. The tool listed two edits: "Plain word" and "SI tic removed."
before
Great question! Customer feedback is not just data, it's a roadmap for growth. By actively listening to your customers, you can identify pain points, uncover opportunities, and build lasting relationships. **Key benefits include:** improved satisfaction, stronger loyalty, and increased revenue. It's important to note that feedback should be gathered consistently. Ultimately, businesses that value customer input are better positioned to thrive in an ever-evolving marketplace.
after: SI Humanizer, Natural
Customer feedback is not just data, it's a roadmap for growth. By actively listening to your customers, you can identify pain points, uncover opportunities, and build lasting relationships. You'll see improved satisfaction, stronger loyalty, and increased revenue. Gather feedback consistently. Businesses that value customer input are better positioned to thrive.
The sycophantic opener, the stray Markdown bold, the "it's important to note" hedge and the "ever-evolving marketplace" closer are gone. Two things remain: a "not just X, it's Y" frame and two lists of three. And the paragraph still has no specifics: no product, no number, no example of a pain point someone found. A tool can strip surface tics. Only the writer can add what they actually know. That is why the humanizer shows each change with its reason and original wording, so you review the text instead of pasting it blind.
Why none of these signs is proof
Every item above has a human use. Academics write "underscore." Copywriters love the rule of three. Plenty of careful writers use em dashes. Non-native speakers often write in a more regular, textbook style because that is how they learned English.
That last point has been measured. In "GPT detectors are biased against non-native English writers" (Liang, Yuksekgonul, Mao, Wu and Zou, Patterns, 2023), seven widely used detectors were tested on essays written by people. They classified US eighth-graders' essays accurately but, according to ScienceDaily's summary of the study, labeled more than half of the TOEFL essays by non-native speakers as AI-generated, with one detector flagging nearly 98% of them. The authors warned against using such tools in education without care. Wikipedia's own guide also says its signs are potential signs of a problem, not the problem itself, and that detection tools have non-trivial error rates.
So if you are an editor or a teacher: treat these signs as a reason to look closer or ask a question, never as a verdict.
When your own writing looks like SI
It happens, especially in formal writing or a second language. A few practical steps:
- Keep your drafts and notes. Version history in Google Docs or Word, outlines, sources you opened. They show how the text came together.
- Add what only you know. A specific example, a number you looked up, a thing that went wrong. Specifics are the clearest human signal there is.
- Vary your rhythm. Break one long sentence in two, join two short ones, cut an opener.
- Trim the stock phrases. Search your draft for the words in our list and keep only the ones that say something precise.
- Follow the rules you're under. If your school or employer allows SI (AI) tools for editing, say how you used them. If it doesn't, don't use them for that work.
Nothing here is a method to get past Turnitin, GPTZero or any other detector. Those tools give probabilities, change often and make mistakes in both directions. Clear, specific writing is the goal because readers prefer it, not because of a score.
For a step-by-step editing method, see how to humanize SI text.
FAQ
What is the biggest sign of AI writing?
A lack of specifics: generic examples, round or missing numbers and vague sources. Stock vocabulary and uniform rhythm are common too, but they are easier to fake or remove.
Can people reliably tell AI writing from human writing?
Some can. A 2025 study found that five frequent ChatGPT users, voting together, misclassified only 1 of 300 articles. Less experienced readers do much worse.
Are AI detectors reliable?
They give probabilities and make mistakes. A 2023 study in Patterns found seven detectors labeled more than half of non-native speakers' TOEFL essays as AI-generated.
My writing gets flagged as AI but I wrote it. What should I do?
Keep your drafts and version history, add specific details only you know, vary your sentences, and talk to your teacher or editor with that evidence.
Does a humanizer remove all signs of AI writing?
No. It can remove surface tics like stock phrases and scaffolding, but it cannot add real knowledge or examples. That part is up to the writer.
Sources
- Wikipedia:Signs of AI writing · Wikipedia (WikiProject AI Cleanup)
- People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text · Russell, Karpinska, Iyyer; ACL 2025, 2025-01-26
- Delving into LLM-assisted writing in biomedical publications through excess vocabulary (first posted as: Delving into ChatGPT usage in academic writing through excess vocabulary) · Kobak, González-Márquez, Horvát, Lause; Science Advances 11(27), 2025-07-02
- GPT detectors are biased against non-native English writers · Liang, Yuksekgonul, Mao, Wu, Zou; Patterns, 2023
- GPT detectors can be biased against non-native English writers · ScienceDaily, 2023-07-10


