How AI Text Detectors Work, and Why They Get It Wrong

By the SAA Tool editor, an IT support engineer · Last reviewed 9 October 2026 · 7 min read

An AI detector gives you a number, and a number looks like evidence. It isn't. Every detector, free or paid, makes an educated guess from patterns in the text, and every one gets some texts wrong in both directions. Here is how the main kinds work, why the mistakes are built in, who pays for them, and how our AI Text Detector reaches its score.

Four ways detectors try to tell

1. Predictability: perplexity and burstiness

A chatbot writes by choosing a likely next word, over and over. A detector can run the text through a language model of its own and measure how predictable each word was; the average surprise is called perplexity, and low perplexity reads as AI-like. Burstiness is the related idea that people vary more, mixing short sentences with long ones and plain words with unexpected ones. In a 2023 study by Liang and colleagues, most of the seven detectors tested were built on the GPT-2 language model.

2. Phrase and style signals

Chatbots lean on stock phrases ("a testament to", "plays a pivotal role", "a rich tapestry"), open sentences with Furthermore and Moreover, and format answers with bold headings. They are cheap to count and easy to explain, but people use them too, especially in formal writing.

3. Trained classifiers

A classifier is a model trained on examples labelled human or AI, which learns its own mix of clues. OpenAI built one and withdrew it on 20 July 2023 "due to its low rate of accuracy". On OpenAI's test set it caught 26% of AI-written text and wrongly flagged 9% of human text, and OpenAI warned that such classifiers are "poorly calibrated outside of their training data".

4. Watermarks

A watermark is added while the text is written, not guessed at afterwards. Google's SynthID Text, for example, adjusts the scores the model uses to pick each next word with a secret pseudorandom function, leaving a pattern a matching detector can test for. It can only find text from models that add the mark, and Google says it is less effective on factual answers and that detection confidence can be "greatly reduced" when text is thoroughly rewritten or translated.

Why every detector gets it wrong both ways

No feature belongs only to AI. People can write predictable, evenly paced prose, and a chatbot can be told to write loosely. Wherever a detector draws its line, some human writing lands on the AI side (false positives) and some AI writing on the human side (false negatives).

False negatives are easy to cause. OpenAI noted that AI text "can be edited to evade the classifier" and that its classifier was very unreliable below 1,000 characters. In Liang's study, asking ChatGPT to rewrite its own college-application essays in more literary language cut detection from up to 100% to no more than 13% across the seven detectors.

False positives add up at scale. The UK's Jisc National Centre for AI works through a hypothetical university: 20,000 students × 8 modules × 3 assessments = 480,000 assessments a year, so even a 1% false positive rate means about 4,800 wrong flags. And the rarer real misuse is, the larger the share of flags that land on honest work. Using the figures OpenAI published for its classifier, imagine 1,000 essays of which 100 were written by AI:

• AI essays flagged: 26% of 100 = 26.
• Human essays flagged: 9% of 900 = 81.
• Flagged essays that are human-written: 81 of 107, about 76%.

The 100-in-1,000 split is an assumption, but the lesson holds for any detector: a flag on its own tells you less than it seems to.

Who gets wrongly flagged

The errors don't fall evenly. In the Liang study, published in the journal Patterns, 91 TOEFL essays by non-native English speakers went through seven detectors:

• the average false positive rate was 61.22%;
• all seven agreed that 18 of the 91 essays (19.78%) were AI-written;
• 89 of the 91 (97.80%) were flagged by at least one detector;
• essays by US eighth-grade students were misclassified at an average of 5.19%.

The flagged essays had lower perplexity: a smaller range of vocabulary is more predictable. When the researchers had ChatGPT improve the TOEFL essays' word choices "to sound more like that of a native speaker", the false positive rate fell to 11.77%. When they had ChatGPT simplify the native speakers' word choices, it rose to 56.65%. The detectors were reacting to vocabulary, not authorship.

OpenAI's guidance for educators names the same risk for students learning English and for "students whose writing was particularly formulaic or concise", and says a detector it trained labelled Shakespeare and the Declaration of Independence as AI-generated.

How our AI Text Detector works

Ours is a pattern checker, not an AI model. It runs in your browser with a fixed list of 118 weighted words and phrases plus a few measurements, and nothing you paste is uploaded. It looks at:

• chatbot-style words and phrases per 100 words;
• sentence length variety (very even lengths count as AI-like);
• sentences that open with linking words such as Additionally, Furthermore or Moreover;
• paragraphs of very similar length;
• chat-style formatting, such as bold headings or bold list items, and long dashes;
• signs of casual typing, which lower the score: lower-case sentence starts, "i" on its own, words like "gonna", and missing apostrophes ("dont").

These combine into a percentage from 1% to 99%. Under 25% reads "Looks human-written", 25% to 59% "Mixed or unclear", and 60% or more "Looks AI-generated". The percentage shows how strongly the text has these patterns, not how much of it an AI wrote. It doesn't measure perplexity or read watermarks, it needs at least 80 words of English, and it warns that under 150 words the score is less reliable. We don't quote an accuracy figure for it.

You can watch it go wrong both ways. Paste this deliberately formulaic paragraph:

Public transport plays an important role in modern cities. It reduces traffic and helps people travel to work and school. Furthermore, buses and trains produce less pollution than private cars. Moreover, they give people who cannot drive a way to reach shops and hospitals. Additionally, a good network can attract new businesses to an area. It is important to note that public transport needs regular investment. Governments should improve services and keep fares low. Overall, public transport offers a wide range of benefits for everyone in the city.

It scores 99%: six chatbot-style phrases, four of eight sentences opening with a linking word, and sentences of much the same length. Swap those phrases for plain wording, keep every fact and stay over 80 words, and it drops to about 12%. A person could have written the first version and a chatbot the second, which is why a score can't settle who wrote something.

How to check a text with our AI Text Detector

1. Open AI Text Detector and paste at least 80 words of English into Your text. Check the length with Word Counter if unsure; 150 words or more is steadier.
2. To see the contrast, click Try an AI-style sample (99%) and Try a human-style sample (9%).
3. Click Check text, or press Ctrl+Enter (⌘+Enter on a Mac).
4. Read the percentage and verdict, then the table under What the score is based on.
5. Scroll down to your text: the most AI-like sentences are highlighted and each chatbot-style phrase is marked.

Using a score responsibly

• Never use it as the only evidence. Jisc says decisions "should never be based solely on AI detection", and the UK's Joint Council for Qualifications says detection tools should form part of a holistic approach.
• Read the reasons, not just the number. Could formal training or English as a second language explain the signals?
• Talk to the writer. Ask how they worked, look at drafts or version history, and have them explain a passage in their own words.
• Don't ask a chatbot. OpenAI says ChatGPT has no knowledge of what is AI-generated, and its answers to "did you write this?" have no basis in fact.
• Follow your organisation's rules. Jisc tells staff they "must not simply find an AI detection tool on the internet and run a student's work through it". That includes ours: use it on your own drafts, and for marking follow your institution's approved process.

Checklist for teachers and editors

• Read the whole piece yourself. At an average 238 words per minute, 1,000 words takes about 4 minutes (Reading Time Calculator).
• Check what the writer was allowed to use: grammar checkers, translation and AI help may be permitted.
• Compare the work with the writer's earlier writing; JCQ lists a difference from a student's usual language style as a possible sign.
• Check that every quote and reference exists. JCQ also lists references that can't be found or verified.
• Record what you checked, and give the writer a chance to respond before any decision.

Spotted a mistake or something out of date? Tell us and we'll fix it.