AI-generated text detector
Analyse a text for statistical patterns typical of LLM-generated content. Heuristic, indicative and 100 % in-browser.
No detector (not even commercial ones like GPTZero or Originality) can state with certainty whether a text is AI or human. OpenAI retired its own detector due to high false-positive rate. This tool gives an indicative signal based on statistical heuristics, useful as a warning but never as proof. Never use a detector to accuse anyone of using AI: false positives are especially harmful in academic contexts.
Computed metrics
How it works
AI-generated text detectors try to identify statistical patterns that distinguish LLM output from human-written text. The best-known heuristics are perplexity (how "surprised" a model would be reading the text) and burstiness (the variability of sentence length). Humans write with more rhythmic variety: short sentences punch, long ones explain, while LLMs tend to produce uniform-length text and smoothly varied vocabulary.
Why every detector fails
Three fundamental reasons: (1) 2024-2026 LLMs can mimic human styles if asked, erasing the detectable patterns; (2) very formal human text (papers, corporate reports) shows the same statistical traits as AI-generated text; (3) a human can lightly edit AI text and break the heuristics. Moreover, false positives are especially frequent with short, technical or non-native writer texts.
What a detector is actually useful for
Treat this kind of tool as a signal, not a verdict: if a text scores very high for AI, it deserves a closer review (ask the author, compare with their previous style, conduct oral interviews). Never use it as sole evidence to penalise or accuse. In educational settings, many universities now officially advise against AI detectors due to their high false-positive rate with non-native English students.
Frequently asked questions
Statistical heuristics (burstiness, sentence length, repetition) work in any language. AI filler phrases are the most frequent Spanish translations of common English template sentences. Accuracy is similar to English, perhaps slightly worse because current LLMs are more optimised for English.
Because commercial detection APIs are paid (and not reliable either: OpenAI retired its own detector in 2023 due to false positives). We prefer to be honest: we give a free heuristic indicator and explain its limits, rather than promising certainty with an API that doesn't have it either.
Lightly editing the text (changing punctuation, breaking long sentences, adding typical language errors) reduces the detectability of almost any heuristic. This is one of the reasons detectors are unreliable in practice: minimal editing noise already breaks the patterns.
Related tools
Other calculators and tools related to this one.
Indicative tool based on statistical heuristics. Does not aim to replace commercial tools nor to issue definitive verdicts. Must NOT be used as evidence in academic, professional or legal contexts. AI detectors have high false-positive rates and can misidentify human text, especially that written by non-native speakers.
