How to tell if text was written by AI
Chatbots can write a fluent essay, review or news story in seconds. People now want to know whether a piece of writing came from a person. There is no reliable test that can tell you. This guide explains what AI text detectors can and cannot do, and what you can check instead.
1. Why it is hard to tell
AI-written text has no built-in mark that proves where it came from. People also mix their own writing with AI suggestions, edit AI drafts, or use AI to translate or tidy up their work. Plenty of human writing is plain and predictable too.
Even AI companies have struggled with this. In 2023 OpenAI withdrew its own AI text classifier, saying it was not accurate enough.
2. What detectors do
Tools such as Copyleaks, GPTZero, Pangram and Turnitin's AI writing indicator look for patterns that are common in text from language models, such as how predictable each next word is. They give a score or a percentage.
That score is an estimate. A result of "90% AI" means the tool is fairly confident the text looks like AI writing. It does not mean 90% of the text was written by AI, and it is not proof.
3. Why detectors get it wrong
- False alarms on human writing. Formal, simple or very polished writing can be flagged. A 2023 Stanford study found detectors often flagged essays by people writing in English as a second language.
- Short texts. A few sentences are not enough for a meaningful result.
- Edited or translated text. Light rewording can hide AI text, and translation tools can make human text look machine-made.
- Different tools disagree. The same text can get very different scores from different detectors.
If you try more than one detector, read each result on its own. Do not add the scores together or take an average. If the tools disagree, the answer is unclear.
4. Check the facts and sources
The most useful checks do not depend on who wrote the text. They test whether it is accurate.
- Look up every reference, book, paper and link. Chatbots sometimes invent sources that sound real but do not exist.
- Search for quotes in quotation marks to see whether the person really said them.
- Check names, dates and figures against a reliable source.
Invented references are strong evidence that something went wrong, whoever wrote the text. They are also a problem in their own right.
5. Signs worth a closer look
These can suggest AI was involved. None of them proves it.
- Chatbot phrases left in, such as "As an AI language model" or "Certainly! Here is your essay".
- Confident statements that are wrong, or details that do not fit the subject.
- Smooth, general text that never quite answers the question asked.
- A sudden change from someone's usual style, vocabulary or level of detail.
- Fake reviews that repeat the same structure and phrases across many accounts.
6. AI writing and plagiarism are different
Plagiarism checkers such as Turnitin and Copyleaks compare text with existing sources and show what matches. That is a different question from whether AI wrote it.
A similarity report describes what matches. It does not show intent. Quotations, common phrases and properly cited material can all produce matches. RELTY's sample report calls this similarity found.
7. If you mark, edit or are accused
If you are a teacher or editor
- Do not act on a detector score alone. Turnitin itself says it cannot rule out false positives completely, and that teachers should use its score to start a conversation with the student.
- Compare the work with the person's earlier writing.
- Ask for drafts, notes or the document's version history.
- Talk to the writer about the work. People who wrote something can usually explain how they did it.
- Follow your school's or publication's written policy on AI.
If you are accused
- Ask which tool was used and what result it gave.
- Share your drafts, notes, sources and version history. Google Docs and Microsoft Word both keep earlier versions.
- Point to the tool's own published limits and to the research on false alarms.
8. Decide what you can say
- Verified. The writer or publisher says how it was written and can back it up, for example with drafts or a clear AI disclosure.
- AI-generated or altered. The writer admits it, or there is clear evidence such as chatbot text left in or invented sources.
- Misleading context. The text is real but presented falsely, for example a quote given to the wrong person.
- Unverified. There is not enough evidence. A detector score on its own leaves text here.
In the news
Recent reports from RELTY's AI news that show why this check matters.
- 12 November 2026 · The Guardian
Caine prize clears its winner of using AI
The Guardian reports the prize found Simbarashe Kundizeza's story The Blue Fighter was written by a person, after a month-long investigation. Questions about AI use can take careful work to settle, and the answer can be no.
- 3 November 2026 · The Guardian
Prosecutor cites a case that does not exist after an AI search
A senior Crown Prosecution Service lawyer cited an invented precedent in a High Court dispute. The Guardian says it adds to dozens of AI-related errors in court proceedings. Looking up every reference catches this kind of mistake.
- 25 October 2026 · The Guardian
Western Australia wants its money back over report with AI-invented references
The state government is seeking a refund from Deloitte for a university merger study with errors that included references invented by AI. Looking up every source is a check anyone can do.
- 24 October 2026 · Xataka
AI detector flags a prize-winning Spanish novel (in Spanish)
The Pangram detector rated a novel that won a prize banning AI use as 99% AI-written. Xataka points out that such detectors are not infallible. A score alone does not settle the question.
- 22 October 2026 · The Guardian
Canadian publisher drops novelist over alleged AI use
Boréal ended its relationship with Thélyson Orélien, saying he used AI to help write his debut novel. He denies it. Cases like this show how much depends on evidence beyond the text itself.
Sources
The pages this guide relies on, checked on 25 October 2026.
- OpenAI quietly shutters its AI text classifier due to low accuracy. Decrypt, July 2023.
- GPT detectors are biased against non-native English writers. Liang, Yuksekgonul, Mao, Wu and Zou, Stanford University, 2023 (arXiv).
- AI writing detection update from Turnitin's chief product officer. Turnitin.
Why AI detectors disagree, and how to read their results
How to spot fake reviews and listings made with AI
What are Content Credentials and how do you check them?
How to tell if an image is AI-generated
Browse the RELTY directory of tools
Links to other organisations are for reference only. RELTY has no partnership with, or endorsement from, any of them.
