You cannot tell for certain. What you can do is notice signs: a shape that is too tidy, no detail that only the writer could know, stock phrases, sentences all the same length, and a tone that does not fit the reader. Several of these together make AI more likely. None of them proves it, because plenty of people write this way, and AI detectors are not reliable either. So never accuse someone on the strength of a hunch or a score.
The signs in the text itself
Shape
- It announces and sums up. The first paragraph restates the question. The last one repeats what was just said.
- Everything comes in threes. Three adjectives, three reasons, three bullet points, whether or not there were three things to say.
- Headings and bullets on something small. A question a person would answer in two sentences comes back as a formatted report.
- Both sides, no view. Every point is balanced by another, and the writer never says what they think.
- An offer at the end. A closing line inviting more questions, on a text that was not a conversation.
Words
- Stock words and phrases. "Moreover", "furthermore", "pivotal", "multifaceted", "it is worth noting", "plays a key role", "I hope this email finds you well". Each is fine alone. A page full of them is a pattern.
- Contrast for show. Sentences that set up a claim only to knock it down, again and again.
- Long dashes where a comma or a full stop would do.
- Hedging. "It may be helpful to consider" where a person would say "do this".
- Feelings named, not shown. In stories: hearts racing, waves of sadness, silences that are deafening.
Rhythm
- Even sentences. Each about the same length, each paragraph about the same size.
- No rough edges. No fragments, no slips, no sentence that runs on because the writer got excited. Flawless grammar in a casual message stands out.
Content
- Nothing specific. No name, no date, no place, no price. The text could be about anyone.
- Details that are wrong. A quote nobody said, a book that does not exist, a link that goes nowhere. A made-up source is one of the strongest signs there is, and it is the one you can check.
- It misses the point. A reply that answers the general topic and not the question that was actually asked.
- The tone is off for the reader. A text to a friend that reads like a letter from a bank.
An example
Here are two thank-you notes after the same favour. This one has the signs:
I hope this message finds you well. I wanted to reach out to express my sincere gratitude for your invaluable support during the recent move. Your help truly made a difference, and I am deeply appreciative.
This one does not:
Thanks again for Saturday. I still can't believe you got the sofa round that corner. I owe you a curry, and probably a new pair of gloves.
The second has a day, a sofa, a corner and a pair of gloves. The first could be sent to anyone about anything. That is the real difference, and notice what it is a difference of: one note is specific and one is generic. A person in a hurry could have written the first. An AI given the right details could have written the second.
Why no sign is proof
- People wrote this way first. Models learnt from formal human writing: school essays, company emails, press releases. Anyone trained to write like that produces the same signs.
- Some writers get suspected more. People writing in a second language, careful students and anyone taught a formal style often write evenly and correctly, and that is exactly what gets called AI.
- Edited text loses the signs. A quick rewrite removes most of them, so their absence tells you nothing.
- Short texts give too little. Two sentences are not enough to judge.
The signs tell you a text is generic. They do not tell you whether a machine or a tired person made it so.
Why AI detectors are unreliable
A detector is itself a model making a guess. It looks at patterns, such as how predictable each word is, and gives back a score. It cannot point to evidence, because there is none in the text to point to.
In practice detectors call human writing AI, call AI writing human, and disagree with each other about the same passage. Plain, correct, formal prose is the most likely to be flagged, which is unfair to exactly the writers listed above. A score is a guess with a number on it. It should never be the reason someone is marked down, refused or accused.
The same goes the other way. No tool can promise that a text will get past a detector, and ours does not.
What to do instead
If you need to know whether someone wrote their own work:
- Talk to them about it. A person who wrote something can explain a choice, expand a point and say where a fact came from.
- Compare it with their other writing. A sudden change in voice is worth a conversation. It is still not proof.
- Check the facts and sources. Look up every quote and reference.
- Ask for the drafts. Notes, an outline and earlier versions show work being done.
- Judge the work itself. If it is generic and says nothing, that is a fair thing to mark, whoever or whatever wrote it.
If it is a message from a friend, ask yourself whether it matters. Someone who used AI to find the words for a hard message may still mean every one of them.
If it is a review or an article online, look for the specifics: did the writer actually use the thing, go to the place, name the people?
If your own writing gets taken for AI
It happens, and mostly to tidy writers. Put in the details only you know, let some sentences be short and some long, and cut the opening and closing paragraphs if you do not miss them. Keep your notes and drafts. How to make AI text sound human covers the rewrite step by step, and it is about reading well to a person, not about fooling a checker. If you were asked for your own words, write your own words.
Hemmingway and sounding like a person
Hemmingway-1 was trained to write like a person in the first place, and in our own blind test, Human-Likeness, it came first of the models we compared. How that test works. If you have a stiff draft, your own or one an AI wrote, the AI humanizer rewrites it to read more naturally while keeping what it says. Read the result and change anything you would not say before you send it.
To test your own eye, Spot the AI shows you two texts written for the same brief, one by a person and one by Hemmingway, and asks which is which.
Try HemmingwayDownload the app
Common questions
Can you tell if something was written by AI?
Not for certain. You can notice signs, such as a very tidy structure, stock phrases, even sentence lengths and no specific detail, and several together make AI more likely. People write that way too, so the signs are a reason to look closer and never proof.
Are AI detectors accurate?
Not enough to rely on. They give a guess as a score, they flag writing by real people, they miss AI text that has been edited, and different detectors disagree about the same passage. A score alone should never be used to accuse anyone.
What words does AI use a lot?
Formal connecting words like "moreover" and "furthermore", grand adjectives like "pivotal" and "multifaceted", and filler such as "it is worth noting". None of them is proof on its own. People use them as well, especially in formal writing.
Can a teacher prove a student used AI?
Not from the text or a detector score alone. What helps is a conversation about the work, the student's notes and drafts, and a comparison with what they have written before. A school's own policy says what counts and what happens next.
How do I stop my writing being flagged as AI?
Nothing guarantees it. Writing with specific details, a mix of sentence lengths and your own opinions reads as a person's, and keeping your notes and drafts lets you show how the work was done.