AI Text Detector

Paste a text and analyse whether it was generated by AI. The analysis is local: your text is never sent to any server.

Analyse text
0 words
0%
AI prob.
—

Signal breakdown
SignalValueIntensityWeight
Marks detected in the text
AI-typical phrases Formal connectors
Analysis limitations: This detector uses statistical linguistic patterns, not a classification model. It is not a definitive detector. Formal academic text written by humans may score high. Heavily edited AI text may score low. Use it as guidance, not as conclusive proof.

How the detector works

Unlike tools such as GPTZero or Copyleaks, this detector does not send your text to any server and does not use a trained classification model. All analysis happens in your browser, in real time, using six statistical signals about the linguistic structure of the text.

The six signals

1. AI-characteristic phrases. Language models overuse certain transitional and filler phrases: "it's worth noting", "delve into", "furthermore", "multifaceted", "in the realm of". The tool detects over fifty of these expressions in both English and Spanish.

2. Sentence length uniformity. Humans naturally mix short and long sentences. LLMs tend to produce sentences of very uniform length. The standard deviation of sentence lengths is measured: a low deviation (under 4 words) is a statistical signal of automatic generation.

3. Vocabulary richness (TTR). The type-token ratio (TTR) measures how many unique words there are relative to the total word count. AI texts tend to have slightly more repetitive vocabulary than spontaneous human writing, especially in longer texts.

4. Formal connector density. LLMs use connectors such as "however", "nevertheless", "therefore", "consequently" more frequently than casual human text. The percentage of words that are formal connectors is measured.

5. Mean sentence length. AI models avoid extremes: they rarely produce 2–3 word sentences or 50-word sentences. The mean sentence length usually falls between 15 and 25 words.

6. Paragraph uniformity. AI text tends to have paragraphs of very similar length. Variance in paragraph length is an additional signal of human authorship.

Important limitations

Statistical analysis does not "understand" text. An essay written by a university professor may score high because it shares formal characteristics with AI text. An AI text thoroughly revised by a human may score low. The detector is a guidance tool, not a definitive arbiter.

For very short texts (under 100 words) the statistical analysis does not have enough data and the results are unreliable. The tool will show a warning in that case.

Frequently asked questions

How reliable is this AI text detector?
The detector analyses statistical linguistic patterns: frequency of AI-typical phrases, sentence length uniformity, vocabulary richness, and connector density. It is not a definitive detector. Accuracy is reasonable for longer texts (over 150 words), but may produce false positives for formal academic writing and false negatives for heavily human-edited AI text.
What patterns does it use to detect AI text?
The analysis combines six signals: (1) presence of overused LLM phrases such as 'it's worth noting', 'delve into' or 'furthermore'; (2) sentence length uniformity (AI models tend to produce sentences of similar length); (3) vocabulary richness via the type-token ratio; (4) formal connector density; (5) mean sentence length; (6) paragraph uniformity. Each signal has a weight in the final score.
Can it detect text from ChatGPT, Claude or Gemini?
The tool analyses patterns common to all large language models (GPT-4, Claude, Gemini, Llama, etc.) because they all share similar biases in text generation. It does not distinguish between specific models. It detects the statistical characteristics of AI-generated text in general, not the digital fingerprint of a specific model.
Does it work with text in Spanish?
Yes. The tool automatically detects the language of the text (Spanish or English) and applies the corresponding phrase list. The statistical analysis (sentence uniformity, TTR, connector density) works equally well in both languages. For texts in other languages, the statistical analysis is applied without the language-specific phrase list.
Why can formally written human text give a false positive?
Academic, legal or technical texts written by humans share some characteristics with AI text: long and uniform sentences, repeated specialised vocabulary and many formal connectors. This is why the detector may score formal human texts highly. The analysis is statistical, not semantic: it does not 'understand' the text, it only measures its metric properties.
What makes a text look like it was written by AI?
LLMs tend to produce text with very uniform sentence lengths (low variance), frequent formal connectors ('however', 'furthermore', 'consequently'), characteristic filler phrases ('it is worth noting', 'it should be mentioned'), slightly repetitive vocabulary and paragraphs of similar length. These patterns emerge from reinforcement learning with human feedback (RLHF), which favours structured and formal responses.
Is this the same as GPTZero or Copyleaks?
No. GPTZero and Copyleaks use classification models trained on millions of labelled texts, which requires server infrastructure. This tool runs entirely in the browser without sending your text to any server, using transparent statistical analysis. The advantage is complete privacy and that the method is explainable: you can see exactly which signals triggered the analysis.

Last updated

Detector updated in 2026. The analysis is completely local: your text never leaves your device.