How does AI detection actually work? If you want to check your text before you submit it, it helps to understand how AI detectors work. The methods range from statistical language analysis through neural classifiers to digital watermarking. None of them is perfect, but together they give a meaningful picture.
This article explains the common detection methods, shows their strengths and their limits, and helps you work out what an AI report actually tells you. Mimir Mentor combines AI detection with plagiarism checking in a single scan, so you can review both kinds of signals before submission.
What methods do AI detectors use?
AI detectors identify AI-generated text by analysing the linguistic patterns typical of machine-written prose. The main approaches fall into four categories: statistical metrics, neural classifiers, stylometry and watermarking. In practice most tools combine several of these methods to improve their hit rate.
| Method | How it works | Strength | Weakness |
|---|---|---|---|
| Perplexity analysis | Measures how predictable the word choice is. AI text has low perplexity, meaning highly predictable word choice. | Works well on unedited AI text | Fails on reworked text; biased against non-native speakers |
| Burstiness analysis | Measures the variation in sentence length and structure. Human writing alternates between short and long sentences; AI text is more uniform. | Complements perplexity by looking at the overall structure | Academic writing naturally has lower burstiness |
| Neural classifiers | Trained models that learn patterns from thousands of examples. They pick up subtle features that individual metrics miss. | Highest accuracy on models they already know | May need updates as writing models and text distributions change |
| Stylometry | Analyses individual writing style: vocabulary, sentence patterns, punctuation. Compares this against known profiles. | Spots breaks in style within a single text | Needs reference texts by the same author |
| Watermarking | Invisible statistical signatures embedded as the text is generated. Detection requires a compatible method and, for some designs, access to a key. | Theoretically very reliable | Only works if the AI provider plays along. Can be weakened considerably by paraphrasing. |
What are perplexity and burstiness?
Perplexity and burstiness are the best-known metrics in AI detection. GPTZero, one of the first AI detectors, used both values as the core of its analysis.
Perplexity measures how “surprised” a language model is by a text. If you write “I went to the bakery and bought a…”, a language model expects words like “loaf” or “roll”. If what follows is “kaleidoscope” instead, perplexity is high. Some AI-generated text has low perplexity, but generation can sample less probable tokens too. Predictability depends on the model, settings and kind of text.
Burstiness describes how much sentence length and complexity vary within a text. People naturally alternate between short, punchy sentences and longer, more involved constructions. AI text, by contrast, holds a much steadier level. A text in which every sentence runs to 15 or 20 words looks suspicious to a detector.
Neither metric is proof on its own. Goethe University Frankfurt stresses that AI detection results may only be used in a supporting role and do not count as evidence of AI use by themselves (Source: Goethe University Frankfurt). A Stanford study also found that several widely used GPT detectors consistently misclassified essays by non-native English writers as AI-generated, while essays by native speakers were identified correctly (Source: Liang et al. (2023)).
How do neural classifiers work?
Neural classifiers go beyond individual metrics. They are trained on large volumes of human and AI-generated text and learn to recognise subtle patterns that are invisible to the human eye.
The advantage: they can weigh up many features at once and so achieve higher accuracy than purely statistical approaches. Mimir Mentor’s AI detector gives a model-based indicator for each flagged passage; a passage-level value is not the proven AI-written share of the whole document.
The challenge: new models and different types of writing can reduce performance. Detectors need ongoing evaluation and may need updates; accuracy does not transfer automatically to every model.
Review your text for AI-related signals
What are digital watermarks for AI text?
Digital watermarks are invisible signatures that an AI model embeds in its word choice as it generates text. The idea comes from research by Kirchenbauer et al. (2023) (Source: Kirchenbauer et al. (2023)). Before each word is generated, part of the vocabulary is randomly placed on a “green list”, and the model is nudged towards those words. The resulting text then contains significantly more green-list words than a human author would produce by chance.
Reading the watermark out requires the key, namely the seed of the random number generator. Without that key the watermark is invisible. That makes the method theoretically very reliable, but it also makes it dependent on the AI provider actively taking part. Google DeepMind has deployed such a scheme, SynthID Text, in its Gemini apps and published it in Nature (Source: Dathathri et al. (2024)). A watermark from one provider, however, says nothing about text generated by another provider’s model.
In practice, watermarking has its limits. Sadasivan et al. (2023) show that recursive paraphrasing, where a text is run through a paraphrasing model several times, can significantly reduce the detection rates of watermarks and other detectors, often with only a slight loss in text quality (Source: Sadasivan et al. (2023)). That is why most commercial detectors rely on a combination of statistical methods instead.
How accurate are AI detectors really?
No AI detector reaches 100% accuracy. Reliability depends on several factors: the AI model used, the length of the text, how much post-editing it has had, and the language.
| Factor | Effect on detection | What you should know |
|---|---|---|
| Text length | Longer texts (500+ words) are detected more reliably than short paragraphs. | Upload the whole text for the check wherever you can, not just individual paragraphs. |
| Post-editing | The more heavily an AI text has been reworked, the harder it is to detect. | Your own phrasing and your own examples make the biggest difference. |
| Language | Detection quality can differ between languages, and not every detector supports every language. | Check which languages a detector supports, and treat results in any language as a signal, not proof. |
| AI model | Newer models (GPT-5, Claude 4) produce more human-like text than older ones. | Detectors have to be updated regularly to keep pace with new models. |
| Type of text | Academic writing is harder to classify, because academic language is formal and uniform by nature. | A low burstiness score is not automatically suspicious in academic writing. |
In February 2026 the Administrative Court of Kassel in Germany set out rules of general application on how the use of AI in assessments can be proven (source: Administrative Court of Kassel). Although the rulings are not yet final, they point to a wider trend: universities and courts alike are steadily developing procedures for establishing whether AI has been used.
What does the result of an AI check mean?
An AI check usually gives you a percentage indicating how much of your text was probably machine-generated. Reading that figure correctly is crucial.
The number is a model output, not proof. The meaning of a percentage depends on the tool: it may describe a classification score or the share of text flagged. Read the provider’s definition before interpreting a result such as 30%. The difference matters, because human writing can show those patterns too.
What you should do: look at each flagged passage individually. Check accuracy, clarity and your actual writing process. Do not change sound writing merely to lower a detector score. Keep drafts and disclose any AI use as required. There is more on this in our article on AI detection and how to check your text.
Revise a passage for clarity and precision
How to check your text responsibly before you submit
A systematic self-check helps you find and fix problem passages before your lecturer reads the text.
- Write it yourself: use AI for brainstorming and structuring, but put the text into your own words.
- Show your own thinking: use subject-specific terminology, your own assessments and concrete examples. Do not alter sound writing just to change a detector score.
- Check before you submit: upload your text to mimir-mentor.com. The scan checks for AI content, plagiarism and unsupported claims all at once.
- Review the flagged passages: check accuracy, sources and your writing process for each one. Fill in missing sources with the literature search.
- Document your use of AI: if you used AI as a tool, record it in an AI use declaration. Our guide shows you how to disclose AI use properly.
Our comparison article explains how Mimir’s AI detection differs from Turnitin and other tools.
Mimir Mentor is GDPR-compliant, with servers in the European Union. Your document is processed for the analysis, deleted after 30 days at the latest and never added to a public database. More than 70,000 students already use the platform.
If you want the practical side as well as the technology, read on in AI detection, AI-written text at university: what is allowed? and how to disclose AI use.
Scientific Formulations in Minutes Seconds
FAQ: common questions about AI detection technology
-
Can AI detectors spot every AI-written text?
No. No detector reaches 100% accuracy. Reliability depends on the AI model, the length of the text and how much post-editing it has had. Heavily reworked AI text with personal examples and an argument of your own is considerably harder to detect.
-
What is the difference between perplexity and burstiness?
Perplexity measures how predictable individual words are: low perplexity means predictable word choice, which is typical of AI. Burstiness measures the variation in sentence length and structure across the whole text. Human writing typically has higher burstiness than AI-generated text.
-
Can AI detectors check texts reliably in languages other than English?
Not every detector supports every language, and detection quality can differ between languages. Turnitin, for example, currently offers its AI writing report for English, Spanish, Japanese and Arabic. Whatever the language, treat an AI flag as a reason to review a passage, not as proof.
-
Is an AI detector result enough to prove academic misconduct?
No. Goethe University Frankfurt stresses in its guidance that AI detection software may only be used in a supporting role and is not sufficient as proof on its own. The University of Hohenheim likewise points out that technical proof of AI use is currently not possible beyond doubt. In 2026 the Administrative Court of Kassel in Germany also set out important rules on what can be proven.
-
What are digital watermarks for AI text?
Watermarks are invisible statistical signatures that an AI model embeds in its word choice as it generates text. The underlying method was described by Kirchenbauer et al. (2023). At the same time, Sadasivan et al. (2023) show that paraphrasing can significantly weaken such watermarks.
-
How can I stop my own writing being flagged as AI by mistake?
No writing style can guarantee that a detector will not flag it. Write clearly, retain drafts and notes, and document any AI assistance. Review false positives in context rather than changing sound writing just to lower a score.
-
How does Mimir’s AI detection differ from GPTZero or Turnitin?
GPTZero uses several signals and trained models; perplexity and burstiness are only part of its approach. Turnitin uses its own neural classifiers and is generally only available through institutions. Whether submitted texts are stored depends on the settings for the individual assignment. Mimir combines several methods, is built for academic writing, is GDPR-compliant and deletes documents after 30 days at the most.
