The short answer
To tell if something is written by AI, look for several signs together: an even sentence rhythm, general claims with no specific detail, a balanced structure that never takes a side, repeated transition phrases, no personal experience, and sources you cannot find. No single sign is proof. People write this way too.
The 12 signs, in short:
- Every sentence has the same rhythm
- General claims with no specifics
- A balanced structure that avoids a position
- The same transition phrases again and again
- Stock words and inflated language
- Repeated sentence formulas and lists of three
- A tidy template with a conclusion that repeats the introduction
- No personal detail or lived experience
- Sources and quotes that cannot be found
- Confident statements that are wrong
- A sudden change from the writer’s usual voice
- Leftover chatbot text or formatting
12 signs of AI writing you can check by eye
Every example below is invented for this guide. None is taken from a real person or a real document.
1. Every sentence has the same rhythm
Read a paragraph aloud. People mix short sentences with long ones. AI text often keeps the same length and the same shape, sentence after sentence, so it sounds smooth but flat.
Invented example: “Remote work offers many benefits for employees. It also creates new challenges for managers. Companies must therefore find a careful balance.”
2. General claims with no specifics
The text says things nobody would disagree with and never gets to a name, a number, a place or a date. Ask yourself what you learned from the paragraph. If the answer is nothing, note it.
Invented example: “Technology has changed the way we communicate in many important ways.” A person who knows the topic would more likely write: “My grandmother now sends voice notes instead of letters.”
3. A balanced structure that avoids a position
Each point is followed by its opposite, and the text ends by saying that both sides matter. Hedging words such as “may”, “can” and “often” appear in almost every sentence.
Invented example: “While social media can bring people together, it can also lead to isolation. Ultimately, the impact depends on how it is used.”
4. The same transition phrases again and again
Look at the first word of each paragraph. A run of “Moreover”, “Furthermore”, “Additionally” and “In conclusion” suggests the text was assembled, not argued.
Invented example: Four paragraphs in a row that begin “Firstly,”, “Moreover,”, “Furthermore,” and “In conclusion,”.
5. Stock words and inflated language
Some words became much more common once AI writing tools arrived. Researchers who studied more than 15 million biomedical abstracts found an abrupt rise in certain style words after AI writing tools appeared. Readers often point to grand, vague words used where a plain one would do. The exact words change as the tools change, so do not rely on a fixed list.
Invented example: “In today’s fast-paced digital landscape, it is crucial to navigate the rich tapestry of online learning.”
6. Repeated sentence formulas and lists of three
The same sentence pattern appears several times on one page. Two common ones are “it is not just X, it is Y” and a list of exactly three items.
Invented example: “Reading is not just a hobby, it is a journey. It builds knowledge, empathy and imagination.”
7. A tidy template with a conclusion that repeats the introduction
A broad opening, three points of equal length and a closing paragraph that says the same thing as the first one. Real drafts are usually uneven, because the writer cares about one point more than the others.
Invented example: An essay that opens with “Climate change is one of the most pressing issues of our time” and ends with “In conclusion, climate change remains one of the most pressing issues of our time.”
8. No personal detail or lived experience
The task asks for a personal view, and the text answers with general statements. There are no names, no small details and nothing only this writer could know.
Invented example: A reflection on a school trip that says “The visit was a valuable learning experience that broadened my perspective” and never mentions where the class went or what happened there.
9. Sources and quotes that cannot be found
AI tools can produce references that look real and are not. In one study of references produced by ChatGPT, 55% of the citations from GPT-3.5 and 18% of those from GPT-4 were fabricated. Those were older models, but the check is still worth doing: search for the title, the author and the quote.
Invented example: “(Hartwell & Osei, 2019, Journal of Applied Learning Futures)”, a reference where no search finds the journal, the article or the authors.
10. Confident statements that are wrong
The text states a wrong date, a wrong name or a wrong figure in a calm, certain tone, with no “I think” or “about”. People make mistakes as well, so check the fact itself before you read anything into it.
Invented example: A book report that describes, in detail, a final chapter the book does not have.
11. A sudden change from the writer’s usual voice
This is one of the most useful signs, and you need earlier writing to see it. Compare the text with emails, earlier essays or past articles by the same person. Look at word choice, sentence length and the mistakes they usually make.
Invented example: A colleague whose emails read “hi, can u send the file thx” sends a report that begins “I trust this message finds you well. Please find herein a comprehensive overview.”
12. Leftover chatbot text or formatting
Sometimes part of the chat is pasted in with the answer: a reply to the prompt, an offer to help further, or a blank left for the user to fill in. Headings and bold bullet labels in a short text that did not need them can point the same way. A pasted reply is the one sign on this list that is hard to explain any other way.
Invented example: An essay that starts “Certainly! Here is a 500-word essay on the causes of the First World War” or contains “[insert your name here]”.
Why no single sign proves anything
These signs describe a style, and a style has no owner. AI tools learned to write from text that people wrote, so everything on the list above already existed in human writing. Three findings from published research are worth knowing.
- Untrained readers do poorly. In a 2021 study, people without training told text from the GPT-3 model apart from human text at random chance level. Short training raised accuracy only up to 55%. The readers also gave contradictory reasons for their answers.
- Practised readers do much better. A 2025 study found that people who often use AI tools for writing were accurate at spotting AI-written articles. The majority vote of five such readers got only 1 of 300 articles wrong. That was a group of experienced readers voting together, not one person reading quickly.
- Some writers are flagged unfairly. A 2023 study found that AI detectors consistently misclassified writing by non-native English writers as AI-generated. The same risk applies to a human reader who treats simple, careful English as a sign of AI.
So your own judgment depends on your experience, and it can be wrong in a way that harms the writer. Count signs, look for several together, and keep the question open until you have more evidence.
When people write this way
Human writing shows the same signs in many ordinary situations:
- Writing in a second language. A smaller vocabulary and safe sentence patterns produce even, predictable text.
- Formal and academic writing. Students are taught the five-paragraph essay, the balanced argument and transition words such as “moreover”.
- Business and legal writing. Templates, hedging and neutral wording are normal.
- Heavily edited text. An editor or a grammar tool removes the quirks that make writing sound personal.
- Short text. Two or three sentences do not show a pattern either way.
- A writer who has improved. A change in voice can come from practice, tutoring or a subject the writer cares about.
Wrong sources are not proof either. People misremember titles and copy references from other papers without reading them. For more on text that is flagged by mistake, see AI detector false positives.
What an AI detector adds
A reader notices style. An AI detector measures something a reader cannot see: how predictable the wording is to a language model, across the whole text. That gives you a second kind of evidence, made the same way for every text you check. Our guide on how AI detectors work explains the method.
The Plagiarism Checker Plus AI detector gives an overall AI score and a likelihood for every sentence. The sentence view is the useful part. It shows whether the AI-like text is spread through the document or sits in one section, and you can compare those sentences with the signs you found by reading. If you want to check text from ChatGPT in particular, use the ChatGPT checker.
There is also no number that is safe or unsafe on its own. See what percentage of AI is acceptable?
What to do with a suspicion
Signs and a score together are a reason to ask questions. They are not a reason to accuse someone.
- Write down what you saw. Note the passages and the signs, so you are not working from a feeling.
- Check what can be checked. Look up the sources and quotes. Compare the text with the writer’s earlier work.
- Ask the writer. A person who wrote the text can usually explain it and show how it was written.
The full process is in how to check for AI writing. Teachers can also read how do teachers check for AI? If your own writing was flagged and you wrote it yourself, see how to prove you didn’t use AI.