Can AI-Written Text Be Detected? What Detectors Can and Can’t Tell You

Teachers, editors and clients increasingly want to know whether a piece of writing came from ChatGPT, Claude, Gemini or a person. Many websites offer “AI detectors” that print a confident percentage. This guide explains how those tools work, why their scores are unreliable, and what you can do instead.

How AI detectors work

Large language models write by choosing likely next words. Detectors try to spot that statistical fingerprint. The most common signals are:

• Perplexity: how predictable each word is to a language model. Very predictable text scores low.
• Burstiness: how much sentence length and structure vary. Uniform sentences look more machine-like.
• Word and phrase repetition, and the use of stock transitions such as “furthermore” or “it is important to note”.
• Classifier models trained on examples of human and AI text, which learn their own patterns.

Why the scores are unreliable

• False positives are common. Clear, formal or formulaic human writing (legal text, lab reports, simple English) often looks “predictable” and gets flagged. Research from Stanford in 2023 found that several popular detectors labelled a large share of essays by non-native English speakers as AI-written.
• False negatives are easy. Light editing, paraphrasing or asking the model to vary its style is often enough to lower the score.
• Short texts can’t be judged. A few sentences do not contain enough signal for any statistical method.
• Even the AI companies struggle. OpenAI withdrew its own AI text classifier in July 2023, citing its low rate of accuracy.
• New models change the patterns. A detector tuned on older models may miss text from newer ones.

Because of this, a detector percentage should never be treated as proof, and it should not be the only basis for a grade, a rejected article or an accusation.

Better ways to check authorship

• Ask about the process. Drafts, notes, version history in Google Docs or Word, and the sources used are much stronger evidence than any score.
• Have a conversation. A writer who did the work can usually explain their argument, choices and sources.
• Check the facts and citations. AI text often contains invented references, wrong dates or confident errors.
• Compare with earlier work by the same person for sudden changes in style or level.
• Set clear rules up front about which kinds of AI help are allowed, so the question becomes “was it disclosed?” rather than “can we prove it?”.

What our Writing Style Analyzer does

Our Writing Style Analyzer shows the same signals detectors rely on (sentence length variety, readability, vocabulary variety and repeated words) without turning them into a fake “% AI” verdict. Use it to edit your own writing or as one piece of context, never as proof. For repeated sentences inside a document, try the Duplicate Sentence Finder.

Does Google penalise AI content?

Google’s published guidance says it rewards helpful, original content however it is produced, and acts against content made mainly to manipulate rankings, including mass-produced pages with little value. In practice that means AI-assisted articles need real expertise, accurate facts and something new to say.

Frequently asked questions

Can any tool prove that text was written by AI?
No. Current tools can only estimate, and their error rates are too high to use as proof.

My own writing was flagged as AI. What should I do?
Share your drafts, notes and version history, and explain your sources. Point to the published limitations of detectors.

Is using AI for writing allowed at school?
It depends on your institution and course. Read the academic integrity policy and ask your teacher when in doubt.