AI content detector

How does this detector work?

Unlike commercial detectors (OriginalityAI, GPTZero) using a neural classifier trained on millions of examples, our tool statistically analyzes your text on 5 measurable criteria. Less precise but 100% transparent: you see exactly which indicators reveal AI-generated content.

Analysis runs in your browser (JavaScript), so your text never leaves your machine. No upload, no storage, no tracking. That's why it's instant: no server round-trip.

The 5 measured signals

Our score combines 5 indicators validated by computational stylometry research:

• Burstiness — coefficient of variation in sentence lengths. Humans write in bursts (short, long, sometimes fragmented). LLMs tend to produce similar-length sentences.
• AI tells — frequency of stereotyped transitional phrases ("it is important to note", "moreover", "let's dive into"). These are over-represented in raw LLM outputs.
• Sentence uniformity — percentage of sentences whose length is close to mean. Human texts typically have 30-45% in this range, AI 55-70%.
• Lexical diversity — Type-Token Ratio (TTR): unique words divided by total. AI repeats terms more, especially in long texts.
• Repeated starters — percentage of sentences starting with the same 2 words. AI often reuses the same openers ("It is", "This means", "However,").

How to humanize an AI text

If your score reveals AI content, here's how to humanize:

1. Radically vary sentence length. A 4-word sentence, then a 28-word one, then a fragment. Break the rhythm.
2. Remove stereotyped transitions: "moreover", "it is important to", "in conclusion".
3. Add personal references, anecdotes, sharp opinions. AI plays it safe — be sharp.
4. Introduce minor intentional irregularities (contractions, casual elisions) to break grammatical perfection.
5. Vary openers: replace repeated "It is" and "This is" with active or interrogative structures.

Tool limits

No AI detector is 100% reliable. Recent studies (2024-2025) show that even commercial detectors (OriginalityAI, GPTZero, Copyleaks) have 60-85% accuracy on short texts, with significant false-positive rates — especially on formal or technical writing.

Our heuristic tool is intentionally honest: it tells you *why* a text seems AI (which indicators reveal it), not a binary verdict. Use it as a flag to review and humanize, not as a judge.

Frequently asked questions