How AI detection works, and why a score isn’t proof

AI detectors are now built into many plagiarism checkers, learning platforms and publishing workflows. They produce a confident-looking percentage, and it’s easy to read that number as a fact. It isn’t. A detector estimates probability from patterns, and understanding how it does that explains both what it can tell you and where it goes wrong.

The basic problem detectors try to solve

Text produced by a language model doesn’t come with a hidden label. Once it’s copied into a document, it’s just words. Any detector therefore has to infer, from the writing itself, how likely it is that a model produced it. There are several ways to attempt that.

Method 1: Statistical signals

Perplexity

Language models generate text by choosing likely next words. Perplexity measures how surprising a piece of text is to a language model. If a model finds every next word highly predictable, the text has low perplexity. Machine-generated text often has lower perplexity than human writing, because it was produced by choosing likely words in the first place.

Burstiness

Burstiness describes how much that predictability, or sentence length and complexity, varies across a document. Human writing tends to be uneven: a long, complex sentence followed by a short one, a surprising word choice here and there. AI output is often more uniform. GPTZero, one of the first widely used detectors, popularised perplexity and burstiness as signals when it launched in early 2023.

The weakness

Plenty of human writing is also predictable and uniform: formal reports, technical instructions, and essays by students taught to follow set structures. Statistical signals overlap between human and machine writing, so they can’t separate the two cleanly.

Method 2: Trained classifiers

Most commercial detectors use machine-learning classifiers: models trained on large collections of text labelled as human-written or AI-generated. The classifier learns whatever features best separate the two groups in its training data, which may include perplexity-style signals, word choices and structure.

Classifiers can be more accurate than simple statistics, but they share an important limitation: they are only as good as their training data. When new AI models are released, when text is edited, or when writing comes from groups under-represented in training (for example, particular dialects or non-native speakers), accuracy can drop in ways that are hard to predict.

Method 3: Watermarking

A different approach is to mark AI text at the moment it’s generated. Researchers have proposed watermarking schemes in which a model subtly favours certain word choices according to a secret pattern that a detector can later check for (see Kirchenbauer et al., “A Watermark for Large Language Models”, 2023). Google has also developed SynthID, which includes a text watermarking method.

Watermarks can be far more reliable than guessing from style, but only for text from a model that applies them, checked by a tool that knows the watermark. They don’t help with text from models that don’t use them, and heavy editing can weaken them.

Method 4: Evidence about the writing process

Some schools and tools look beyond the finished text. Version history in Google Docs or Microsoft Word, drafts, notes and writing-replay features show how a document came together over time. This kind of evidence says much more about authorship than any single score, which is why many educators treat it as more meaningful than detector output.

What a detector score actually means

A result such as “78% AI” is rarely a statement that 78% of the words were written by AI. Depending on the tool, it may be the model’s confidence that the whole document is AI-generated, or the share of segments that crossed a threshold. Tools also set their own thresholds for labels like “likely AI”, so the same text can get quite different results from different detectors.

Read the documentation for the specific tool, and look at which passages were flagged rather than only the headline number.

Why detectors make mistakes

Human and AI writing overlap

Simple, formal or formulaic human writing can look statistically like AI text. Heavily edited AI text can look human.

Some writers are flagged more often

A 2023 study by researchers at Stanford tested several popular GPT detectors on essays written by non-native English speakers for the TOEFL exam. The detectors misclassified more than half of those essays as AI-generated, while performing far better on essays by US eighth-graders (Liang et al., “GPT detectors are biased against non-native English writers”, published in Patterns). The authors linked this to lower perplexity in writing that uses a more limited vocabulary.

Short texts carry little signal

A paragraph or two gives a detector very little to measure, so results on short passages are much less stable.

Even the builders have struggled

OpenAI released an AI text classifier in January 2023, noting that in its own evaluations it correctly flagged only 26% of AI-written text while wrongly labelling human text as AI 9% of the time. In July 2023 the company withdrew it, citing its low rate of accuracy. That same summer, Vanderbilt University announced it was disabling Turnitin’s AI detection tool, citing concerns including false positives and a lack of transparency about how it worked.

Small error rates add up

Even a low false positive rate matters at scale. Here’s a hypothetical to illustrate: if a detector wrongly flagged just 1% of human-written essays, a university checking 50,000 genuine essays a term would flag around 500 students who did nothing wrong. Turnitin itself advises that its AI indicator should not be used as the sole basis for action against a student.

Using detectors responsibly

If you’re an educator or editor: treat a score as a reason to look more closely, never as proof. Compare the writing with the person’s other work, look at drafts and version history, and talk with them before drawing conclusions.

If you’re a writer: keep your drafts and notes, follow the AI rules that apply to you, and disclose AI use when required. If you’re wrongly flagged, our guide to AI detector false positives explains how to respond.

You can see how pattern-based signals behave on your own text with our AI detector. It explains which signals it found and how confident it is.