Updated November 21, 2026. There's no reliable way to prove a piece of text was written by AI. AI detectors can be wrong in both directions: they miss AI text and flag human writing as AI. The best approach is to look for several warning signs together, compare the text with the writer's usual work, check facts and sources, and talk to the writer, rather than relying on a detector score.
This guide is for teachers, editors, employers and anyone who wants to judge text fairly.
Why detecting AI text is so hard
AI language models are trained to produce fluent, natural-sounding writing. The better they get, the more their output looks like ordinary human writing. And people edit AI drafts, mix them with their own words, or ask the AI to write in a particular style, which blurs the line further.
Even OpenAI, the company behind ChatGPT, couldn't build a reliable detector. In January 2023 it launched an "AI classifier", and its own announcement said that in tests it correctly identified only 26% of AI-written text as "likely AI-written", while wrongly labeling human-written text as AI-written 9% of the time. OpenAI listed further limits: the tool was unreliable on text under 1,000 characters, and AI-written text could be edited to evade it.
In July 2023, OpenAI withdrew the tool. The same page now says that as of July 20, 2023, the classifier is no longer available due to its low rate of accuracy. TechCrunch and Times Higher Education both covered the decision.
Do AI detectors work?
Many AI detectors are available, some free and some paid. They can sometimes give a useful signal, but you should treat any score as a hint, not proof. Here's why:
False positives: detectors can flag human writing as AI, especially formal, simple or formulaic writing. Writers using a second language may be more likely to be flagged because their writing can be more predictable.
False negatives: lightly edited or paraphrased AI text often passes as human.
Short texts: OpenAI found its classifier was unreliable under 1,000 characters, and short texts are hard for any detector.
Changing models: detectors trained on older AI models may struggle with newer ones.
No ground truth: a detector can't show you how it reached its score in a way you can verify.
Never punish someone based on a detector score alone. A wrong accusation can seriously harm a student, job applicant or writer.

10 signs text may have been written by AI
None of these proves anything on its own. Look for several together, and compare with the writer's other work.
Generic, polished but empty content: lots of smooth sentences that say little.
Repetitive structure: every paragraph the same length, often with a neat summary sentence at the end.
Overused phrases: such as "in today's fast-paced world", "it's important to note", "delve into" and "a testament to".
Lists of three everywhere: "efficient, effective and engaging".
No personal detail: no specific memories, examples or opinions where you'd expect them.
Confident errors: made-up facts, wrong dates or quotes that don't exist.
Fake or broken sources: references to papers, books or links that can't be found.
Sudden change in style: writing that's very different from the person's previous work.
Off-topic padding: text that answers a slightly different question than the one asked.
Inconsistent spelling or tone: such as American spelling in a British student's essay, or sudden formality.
A fair process for checking suspected AI writing
For teachers and lecturers
Check your policy. Know what your school or university allows. Many now permit some AI use if it's declared.
Compare with earlier work written in class or under supervision.
Check the sources. Do the references exist? Do they say what the essay claims?
Talk to the student. Ask them to explain their argument, how they found their sources and how they wrote it. Someone who wrote the work can usually discuss it easily.
Look at drafts and version history, for example in Google Docs or Microsoft Word, if available.
Design work that's harder to outsource: personal reflection, in-class writing, oral presentations and assignments based on class discussions.
For editors and publishers
Ask writers to declare AI use, and set a clear policy on what's acceptable.
Fact-check claims, quotes and statistics; AI-generated errors are often the strongest evidence.
Ask for notes, interview recordings or research materials behind the piece.
For employers and recruiters
Judge applications on substance, not on a detector score.
Use interviews and practical tasks to check skills claimed in writing.
Be clear in job postings about whether AI-assisted applications are acceptable.

A fair process in four steps:
Step | What to do |
1. Look for evidence | Compare with the writer's earlier work and drafts |
2. Don't rely on detectors | Treat any score as a clue, never as proof |
3. Talk to the writer | Ask them to explain their ideas and sources |
4. Decide fairly | Follow your school's or employer's published policy |
What about watermarks and labels?
Some AI companies are working on ways to label or watermark AI-generated content, especially images, audio and video. For text, watermarking is harder, because a few edits can remove the signal. When OpenAI withdrew its classifier, it said it was researching more effective provenance techniques for text. For now, don't assume that text without a label was written by a human.
If you're a writer: how to avoid being falsely flagged
If you write your own work, you can still protect yourself:
Keep drafts and notes, or write in a tool with version history.
Use your own examples and experiences. Specific details are hard to fake.
Declare any AI help your school or employer allows, such as grammar checking or brainstorming.
Don't run your work through "AI humanizer" tools, which can damage your writing and may break rules.
If you use AI to help you write, our guide to writing AI prompts explains how to use it as an assistant rather than a ghostwriter.
Frequently asked questions
Is there a 100% accurate AI detector?
No. OpenAI itself said it is impossible to reliably detect all AI-written text, and withdrew its own classifier because of low accuracy.
Can ChatGPT tell if it wrote something?
No. ChatGPT can't reliably recognize its own past output, and asking it "did you write this?" can give a confident but wrong answer.
Are free AI detectors worth using?
They can give a rough signal, but treat results with caution. Use them alongside other checks, never as the only evidence.
Is using AI to write cheating?
It depends on the rules where you are. Many schools and employers allow some AI use if you declare it. Always check the policy first.
3 tips before you go
If you're a student, keep your drafts and notes; version history is the best defense against a false accusation.
If you're a teacher or manager, ask for a short conversation about the work before drawing any conclusion.
Check whether facts and quotes in the text can be traced to real sources; invented details are a stronger signal than style.
