How Do AI Detectors Work, and How Accurate Are They?

AI detectors do not read minds or find a watermark. They measure how predictable a text is, or they use a classifier trained on examples of human and AI writing, then output a probability-style score. That score is an estimate, not proof. Errors are real, especially for non-native writers, formulaic text and edited AI output, so a score should never be the only evidence against a person.

The short answer

An AI detector is a statistical tool. It looks at patterns in a piece of text, such as how predictable the word choices are, and compares them with patterns it learned from samples of human and machine writing. It then returns a score. The score says "this reads like the AI samples" or "this reads like the human samples". It does not say who wrote the text.

That is why every detector can be wrong in both directions. It can label human writing as AI (a false positive), and it can miss AI writing that was edited or rewritten (a false negative). Treat any result as a reason to look more closely, not as a verdict.

The two main approaches

Most detectors use one of two ideas, and many combine them.

1. Statistical signals: perplexity and burstiness

A language model predicts the next word in a sentence. Text produced by such a model tends to use words the model itself finds likely. Two measures follow from this:

• Perplexity describes how surprising a text is to a reference model. Low perplexity means the wording was easy to predict. Human writing often contains more unusual choices, so on average it scores higher.
• Burstiness describes how much sentence length and complexity vary. People tend to mix short and long sentences. Some machine text is more even.

The weakness is obvious once you think about it. A careful writer who uses plain, standard phrasing also produces predictable text. A famous passage that appears often in a model's training data looks "predictable" for the same reason. Predictability tells you about the text, not about its author.

2. Trained classifiers

A classifier is a model trained on many labeled examples: texts known to be human, texts known to be generated. During training it learns which features separate the two groups. Those features can be word and character patterns, punctuation habits, sentence rhythm, vocabulary range and many others. At test time it reads your text and outputs a probability.

Classifiers depend heavily on their training data. If the examples came from a few older AI systems, the classifier may spot those systems well and miss newer ones. If the human examples were mostly one kind of writing, say news articles, it may judge other kinds of human writing more harshly.

Other ideas you may hear about

• Watermarking means a generator deliberately biases its word choices in a way that a key holder can later detect. It only works if the generator applies it and the text is not rewritten, and it does not help with text from systems that do not use it.
• Provenance records such as version history in a document editor show how a text was produced. They are evidence of process, which a detector score is not.

Why a score is an estimate

A detector score is calibrated on test data. A result such as "80%" does not mean there is an 80% chance that this particular text is machine-written. It means the text resembles the machine examples more than the human ones, on the scale the model was built with. The score also depends on:

• Length. Short texts give the model little to measure, so scores on a few sentences are much less reliable than scores on several pages.
• Genre. Lab reports, legal clauses, product descriptions and cover letters follow templates. Template-driven human writing can look machine-like.
• Editing. AI text that has been paraphrased, edited or mixed with human passages often drifts into the uncertain middle or looks human.
• Language. Detectors are usually built and tested mostly on English. Accuracy in other languages is often lower, and an honest tool should say whether it was measured for that language.

In 2023 OpenAI withdrew its own AI text classifier, citing its low rate of accuracy. The lesson is not that detection is useless, but that the technique has limits and that vendors should publish them.

False positives: who gets hurt

The most serious problem is a human writer flagged as a machine. Researchers at Stanford published a study in 2023 (Liang and colleagues, in the journal Patterns) reporting that several widely used detectors misclassified writing by non-native English speakers more often than writing by native speakers. The likely reason is the one above: a smaller active vocabulary and standard phrasing produce lower perplexity.

Others at risk include:

• Students writing to a strict template, such as a five-paragraph essay or a lab report.
• Writers using grammar tools heavily. A spell checker is harmless, but tools that rewrite whole sentences make text more uniform.
• Technical and legal writers, whose conventions limit variety on purpose.
• Writers who revise toward "clean" prose. Plain, correct English can read as machine-like.

An example. This human-written sentence from a methods section is typical of scientific English:

The samples were centrifuged at 4,000 rpm for ten minutes and the supernatant was discarded.

Every word is expected. A detector measuring predictability may score it as machine-like, yet nothing about it suggests who wrote it. A sentence full of personal detail behaves differently:

By the third week of fieldwork Mrs. Okafor was greeting us in Igbo, which our interpreter said was a mercy to everyone.

Specific, unusual detail is harder for any model to predict. But even this is no guarantee. A person can write bland text, and a machine can be prompted to write vivid text.

How AdeptWrite's AI Content Detector works

The AI Content Detector uses a hashed n-gram and stylometric model. In plain terms:

1. It breaks your text into overlapping character sequences and short word sequences (n-grams), converts them into numeric features with a fixed hashing function, and adds a small set of style measures such as sentence-length patterns.
2. A trained model turns those features into a score.
3. You get an overall band and a score marked as approximate ("≈"), plus highlighted sentences that read most machine-like.

The statistical scoring runs in your browser. An optional "deep analysis" button sends your text to an AI model on Cloudflare Workers AI for a second opinion, for that request only. The tool takes up to 6,000 characters per check, and it needs a minimum amount of text (about 40 words) before it will score anything. It tells you when a text is too short.

What the result means, in the tool's own terms:

• The score is an estimate, not proof. It is calibrated on test data with equal numbers of human and AI texts. It is not the chance that your text was written by AI.
• The highlighted sentences are weaker evidence than the overall score.
• AI text that has been paraphrased, edited or "humanized" is usually missed.
• Simple, formulaic or non-native writing can be wrongly flagged.
• Documents with both human and AI passages often land in "mixed or uncertain".
• New AI models write differently, so the detector may miss them.
• Text corrected or rephrased with AI writing tools, including AdeptWrite's own grammar checker and paraphraser, can be labeled mixed or likely AI.

The tool page publishes the measured test results and the per-language figures in its methodology section, generated from the shipped model so that the numbers cannot drift. We do not quote a single "accuracy" figure here, because one number hides the trade-off between catching AI text and wrongly flagging people. Read the page's measured true-positive and false-positive rates together.

How to read a score responsibly

Use this as a checklist whether you are a writer, a teacher or an editor.

Better evidence than a score includes version history, saved drafts, notes, sources the writer can discuss and a short conversation about the argument. A student who can explain every choice in an essay is demonstrating authorship in a way no detector can.

What to do if you are flagged

If a detector flags your own writing:

1. Stay calm and ask what the evidence is. A score alone should not decide a case; the AdeptWrite tool's own result says so.
2. Gather process evidence: document history, outlines, notes, drafts, search history, sources you consulted.
3. Offer to explain your work. Talk through your argument, sources and word choices.
4. Run the same text through more than one tool. Disagreement between tools shows how uncertain the signal is.
5. Ask for the institution's written policy on detectors and appeals. Policies differ between schools, universities and employers, so read yours rather than assuming.

Our guide on being falsely accused of using AI goes through this in more detail.

If you use AI tools in your own writing

Rules on AI assistance vary by course, journal and employer. Some allow it with disclosure, some forbid it. Check the policy first, then keep your drafts. If you are working from AI output, rewriting it substantially in your own voice and checking every claim is both better writing and a better defense. See how to humanize AI text for editing steps, and is using AI plagiarism? for how institutions treat it.

Plagiarism checking is a different question from AI detection. A plagiarism checker compares text with existing sources; an AI detector estimates style. A text can be fully original and still read as machine-like, or copied from a human source and read as human. Use the Plagiarism Checker for matching sources, and see how to check for plagiarism for the process.

Summary

• Detectors estimate; they do not prove.
• The two main methods are statistical signals (perplexity, burstiness) and trained classifiers.
• Short, formulaic, edited and non-native writing are the hardest cases.
• Publish limits, measure per language, and never use one score as the sole evidence against a person.
• Process evidence, drafts and a conversation are stronger than any number.

← All guides