How AI detection works
Detectors measure two signals: perplexity (how predictable each word is given the previous words) and burstiness (how much sentence length varies). AI-generated text is statistically more predictable and more uniform than human writing.
Modern detectors combine these signals with a classifier trained on millions of paired human / AI samples.
Step-by-step
- 1Open the AI Detector.
- 2Paste the text you want to analyse — at least 200 words for reliable scoring.
- 3Click Detect.
- 4Read the AI-probability score and the per-sentence breakdown.
- 5Use the score as a signal, not a verdict — review flagged sentences manually.
When to use a detector
- Editors triaging freelance submissions for fact-check effort.
- Teachers spotting essays that may need a follow-up conversation with the student.
- Hiring managers vetting take-home assignments.
- Marketing leads making sure agency deliverables aren't entirely AI-written.
Why no detector is 100% accurate
Heavily edited AI text starts to look human. Carefully written human text by someone with a clean, simple style can score as AI. False positives and false negatives both exist.
Use detector output the way you'd use a spam filter — as a triage signal, not a courtroom verdict.
What detectors actually measure
AI text detectors do not read your document and understand it. They measure statistical properties — perplexity (how predictable each word is, given the previous ones) and burstiness (how much sentence length and complexity varies). Human writers spike unpredictable word choices and varied rhythm; large language models, optimised to minimise perplexity, produce smoother and more uniform output. A detector flags passages that sit in the LLM-like region of those metrics. That's a useful signal but it is a signal, not a verdict.
Why false positives happen
- Non-native English speakers often write with shorter, more uniform sentences — the same fingerprint detectors associate with AI.
- Highly-edited prose (academic papers, polished marketing copy) flattens burstiness because editors smooth out idiosyncrasy.
- Templated formats — cover letters, product descriptions, recipe intros — naturally repeat phrasing and trigger false positives.
- Short passages (under 200 words) don't give the model enough signal; results are noisy.
- Translation from another language tends to produce LLM-like prose because translators favour clarity over voice.
How to use detector results responsibly
Treat a high AI-score as a prompt for conversation, not as proof. In an academic context, ask the writer to walk you through the drafting process, show earlier versions, or explain a specific paragraph in their own words. In a hiring context, replace take-home essays with live writing tasks if AI authorship is a concern. In a publishing context, focus on whether the content is accurate, original, and useful — those matter more than which tool typed the first draft. Detectors are diagnostics, not disciplinary tools.
FAQ
- Is the AI Detector free?
- Yes — paste any text and detect for free.
- What text length works best?
- 200+ words. Short snippets are inherently hard to classify.
- Does it work on non-English text?
- It works best on English. Accuracy drops in other languages.
- Can I bypass detection by paraphrasing?
- Sometimes. Mixing manual rewrites with paraphrasing reduces detection probability but never eliminates risk.
- Is my text stored?
- No — input is analysed and discarded.
- Should I trust a single score?
- No. Combine detector output with editorial judgement and, for academia, with a plagiarism check.