Sign in or create an account

How AI detection works

A detector looks at a piece of writing and estimates whether it reads more like a person or more like a language model. Here is what that involves, and why the answer is a signal rather than a verdict.

Updated

The short version

An AI detector is a classifier. It is a model trained on large collections of writing that people produced and writing that language models produced. After training, it can look at new text and say which group that text most resembles. It does not look the text up anywhere, and it cannot see how the text was made. It only sees the words on the page.

That has one important consequence: a detector gives you a resemblance, not a fact. A result tells you how a passage compares with patterns the model learned. It cannot tell you who typed the words, or whether someone used a tool along the way.

The main ideas in plain language

  • Perplexity. A measure of how surprised a language model is by each next word. Predictable text scores low and unusual wording scores high. A 2023 study of GPT detectors says most of them use perplexity, and warns that it can unfairly flag writers with a narrower range of vocabulary and grammar ( Liang and colleagues, Patterns, 2023 ).
  • Burstiness. How much that predictability varies from sentence to sentence. GPTZero described it in a 2023 post as a signal where steady, uniform patterns point to AI, and says it later moved to a deep-learning model ( GPTZero, March 2023 ).
  • Classifier models. A model trained directly on labelled examples of human and machine writing, which learns its own signals instead of relying on one hand-picked measure. This is the general approach described above.
  • Watermarking. Not detection of style at all: a model's maker adjusts its word choices so the text carries a hidden statistical signature that only the maker's own detector can check. It says nothing about text from other models. See what Google's SynthID does and does not cover.

What happens when you check a text

On this site the steps are simple. You paste text or upload a file, and we pull out the plain text. We send it to the detection provider, Pangram Labs, which splits it into passages and labels each one. We then turn those labels into the result you see: a verdict, a share of the text that was flagged, a confidence level, and highlights on the individual passages.

Pangram labels each passage as AI-Generated, AI-Assisted or Human Written, and attaches a confidence of High, Medium or Low. The detector returns labels. It does not explain why a passage got its label, and we do not make up reasons on its behalf. If a highlighted sentence looks perfectly ordinary to you, that is a normal outcome and a good reason to read the surrounding text before drawing a conclusion.

Why longer text is easier to judge

A single sentence carries very little information about its author. Plenty of people write short, tidy sentences, and language models can write short, messy ones. Over several paragraphs, more evidence builds up and the label becomes steadier. That is why this checker needs at least 50 words to run at all, and why we suggest 150 words or more for a result worth reading. The methodology page covers the limits in full.

What a "share" means

Because each passage is labelled separately, the result is naturally a proportion. If 30% of the text is flagged as AI-like, that means about three tenths of the passages were labelled that way. It does not mean there is a 30% chance that the whole text was written by AI, and it does not mean 30% of the words were typed into a chatbot. People often misread this number, so we call it "Flagged as AI-like" and keep it separate from the confidence level.

Where detection goes wrong

Any classifier makes two kinds of mistakes. A false positive labels human writing as AI-like. A false negative lets AI-influenced writing pass as human. Both happen. Some situations make them more likely:

  • Short passages, which give the model little to go on.
  • Heavily edited text, where a person has reworked AI output or an AI tool has polished a person's draft.
  • Translated text, because translation changes the surface patterns of the original.
  • Formal, templated or very careful writing, such as legal boilerplate or a polished cover letter.
  • Writing by non-native speakers, which can follow patterns the model has seen less often.
  • Languages other than English, where detection is generally less reliable.

Treat a result as one input to a conversation. If it matters, look at drafts, version history and the person's other writing, and ask them about their process. Nobody should be penalised on a detector result alone.

General methods and what runs on this site

Detectors across the industry differ. Some score how predictable the wording is, some train a classifier directly on labelled examples, and some combine approaches. Vendors rarely publish full details, so be wary of any page that claims to explain exactly how a given commercial tool decides.

On this site the only thing we can state for certain is the pipeline above: your text goes to Pangram Labs, which returns a label and a confidence for each passage. We do not run our own detection model, and we do not know the internals of Pangram's model beyond what it publishes. Everything we add is presentation: turning those labels into a share, a verdict and highlights.

Other languages and how detectors get evaluated

Most published detector testing is in English. Results for other languages, mixed-language text or translated text should be treated as less certain, even where a vendor lists support. When researchers evaluate a detector, they test it on sets of texts with known authors and count how often it is right, and how often it wrongly flags human writing. Results depend heavily on the texts chosen, their length and the date, so a single accuracy figure says little. We publish none of our own; the accuracy page explains why.

Detectors are not plagiarism checkers

A plagiarism checker compares a document with existing sources and reports matches. An AI detector does something different. It does not compare your text with any database, so freshly written AI text can score as AI-like even though it has never appeared anywhere else, and copied human writing can score as human. The two tools answer different questions and do not replace each other.

How to use the result well

  1. Check enough text. A full essay or a few paragraphs tells you far more than one sentence.
  2. Read the highlights, not only the headline number. Look at which passages were flagged and whether they read differently from the rest.
  3. Pay attention to confidence. A Low-confidence label is a weak signal.
  4. Consider the context: who wrote it, in which language, under what conditions, and whether any editing tools were allowed.
  5. Gather other evidence before acting on a concern.

When you are ready, run a check. If you want to know what we can and cannot say about reliability, read our accuracy page.