How AI Detection Tools Work, and Why They Get It Wrong

AI detectors are programs that try to guess whether a piece of writing came from a person or a language model. They do not read minds and they do not check a database of machine-written sentences. They measure statistical quirks and return a score. That score is a guess, and it is wrong often enough to have wrecked real people's results.

What a detector actually measures

A detector pushes your text through a language model of its own and looks at how predictable each word was. Human writing tends to mix rare and common words and varies sentence length a lot; machine text often sits in a narrower band. Two measures come up repeatedly. Perplexity asks how surprised a model is by each word choice. Burstiness asks how much sentence length and complexity vary across a passage.

Some tools add a classifier, which is a model trained on a labelled set of human and machine samples. Others look at punctuation habits, spacing, or the frequency of stock phrases. None of them verify authorship. They compare style against a pattern and print a likelihood.

Why false positives happen

A false positive means human writing flagged as machine writing, and it is common. The people hit hardest are often those writing in a second language, because taught patterns produce a narrower vocabulary and tidier sentence structure, which is exactly what the statistics reward. Poets, lawyers and technical writers who repeat structures get flagged too.

Short samples make it worse. A few hundred words give a detector very little to work with, and scores swing wildly between paragraphs of the same essay. Quoted passages, references and pasted definitions can be flagged even though the student did not write them. Detector companies publish their own accuracy figures, and many of those figures place the error rate high enough that one flagged paragraph means almost nothing on its own.

Why a score is not proof

A detector output is a probability, not a finding. There is no forensic chain linking a document to a particular model. The same essay can be scored by two tools and land at opposite ends of the scale. Re-running one tool on the same text can shift the result again.

That has practical consequences. Misconduct cases built mainly on detector scores have collapsed on appeal, and journalists have documented students penalised for work they wrote themselves. When a school or an employer treats a score as decisive, the person accused is left trying to prove a negative, which is close to impossible. Detector output should be one weak signal among several, and it should never be the one that decides.

Why detectors struggle to keep up

Detectors are trained on the models of yesterday. When a new model ships, or when someone runs text through a paraphrasing step, or edits it by hand, the statistical fingerprints fade and accuracy drops. Buying a different tool does not fix this, because the problem is structural rather than a matter of tuning.

Watermarking, which embeds a hidden signal in generated text, is a more promising direction, but it requires the model maker to cooperate and it does not survive translation or heavy rewriting. Researchers have also shown that inserting a few deliberately odd sentences can push a flagged essay back under the threshold without changing what it means.

Words to know

• Perplexity - How surprised a model is by each word. Low perplexity means text that was easy to predict.
• Burstiness - How much sentence length and structure vary within a passage. Human writing usually varies more.
• False positive - Human writing that a tool labels as machine writing.
• False negative - Machine writing that a tool passes as human.
• Classifier - A model trained on labelled examples to sort text into categories.
• Watermarking - A hidden pattern inserted into generated text so it can be identified later.

If you are accused of AI use

Ask what evidence exists beyond the score. Request the tool name, the exact passage that triggered it, the version used and any stated error rate. Offer your drafts, notes, reading list and revision history, because that trail is far stronger than a number printed by software. Ask for a conversation about the work. A short discussion of your own argument is something a detector cannot fake and a ghostwriter cannot survive.

Whatever the outcome, make your process visible from then on. Save dated drafts, keep your research notes, and write in a document that records version history. The best defence was never a better detector. It is a paper trail built while you worked.