Are AI detectors accurate?
What a score means for your paper
AI detectors do not know who wrote a text. They estimate how much it reads like the output of a language model, and that estimate is wrong often enough that no detector score proves anything on its own. Here is what the score measures, where it fails, and what to do before you submit or if your own writing is flagged.
What an AI detector actually measures
A detector reads a passage and rates how closely it matches the habits of machine writing: stock phrases, generic claims, a very even sentence rhythm, summaries that restate rather than add. It does not see your drafts, your notes or your writing history, and it cannot trace text back to a particular tool.
The result is a probability-like score per passage, usually rolled up into a percentage for the whole document. That number describes the text, not the writer.
How often detectors get it wrong
Two kinds of error matter. A false positive flags human writing as AI-written; a false negative misses AI text. Every detector makes both, and the balance depends on where its threshold sits.
In our own tests on papers written before ChatGPT existed, about 1 human passage in 40 was flagged, and about 7 in 10 passages written by AI models were caught. Studies of commercial detectors report the same pattern: they flag some human writing and miss AI text that has been edited.
Whose writing gets flagged
- Formulaic sections: methods, standard definitions and boilerplate such as ethics or data statements read alike in every paper.
- Writing by non-native English speakers, which tends to use common phrasings and a regular structure.
- Text heavily polished with grammar tools, which smooth out the irregularities of human writing.
- Very short passages, where there is too little text to judge either way.
Why edited AI text slips through
Detectors look for the surface habits of model output. Once a person rewrites sentences, adds specific details and reorders the argument, those habits fade, and so does the score. Text a model only polished, rather than wrote, is rarely caught at all.
This is why a low score does not certify that no AI was used, and a high one does not prove that it was.
If your own writing is flagged
- Stay factual
A detector score is not evidence of misconduct, and many universities say so in their guidance. Ask which passages were flagged and what tool produced the score.
- Show your process
Draft versions with dates (Word or Google Docs version history, Overleaf history, Git), notes, outlines and the sources you read are the strongest evidence of authorship.
- Explain the flagged passages
Methods text, definitions and phrasing you reused from your own earlier work are common triggers. Point to where each came from.
- Check your institution’s policy
Many allow AI for language editing but not for content. Declare any use the policy asks you to declare.
Using a detector on your own draft
Run before you submit, a detector is most useful as a reading guide. The passages it flags are often the generic ones: sentences that could appear in any paper on the topic. Rewriting them with your specific data, your reasoning and your field’s terms makes the paper better whatever the score.
Do not chase a score of zero by paraphrasing tools or synonym swaps. That makes the writing worse and does nothing for the argument.
Questions
Can an AI detector prove I used ChatGPT?
No. A detector estimates how much a text reads like machine writing. It cannot identify the tool or the writer, and it flags some human writing, so its score alone proves nothing.
Why was my human-written paper flagged as AI?
Formulaic sections, common academic phrasing, grammar-tool polishing and non-native English all make text read more like model output. Ask which passages were flagged and show your drafts and notes.
What percentage of AI is acceptable in a thesis?
There is no universal threshold. Your institution’s policy decides what use of AI is allowed and whether it must be declared; a detector percentage does not.
Do AI detectors work on PDFs?
Some do. The NoveltyAxis AI Detector reads the text of a PDF page by page, leaves out headings, captions, tables and the reference list, and shows each flagged passage with its page.