How AI text detectors actually work
There is no fingerprint
The intuition most people have is that AI-generated text carries a hidden marker a detector can find. It does not. A language model produces ordinary words in ordinary order; nothing is embedded, nothing is signed. What detectors do instead is measure statistical properties of the text and compare them against what human and machine writing typically look like. That is a fundamentally different and much weaker operation, and every strength and failure of detection follows from it.
Perplexity: how surprising is the next word?
Perplexity measures how well a language model predicts a text. Feed it a passage and ask, at each word, how probable that word was given everything before it. Low perplexity means the text went where the model expected; high perplexity means it kept surprising the model. This matters because generating text is, mechanically, choosing likely next words. A model asked to write about climate policy will tend toward the phrasings that are statistically common in writing about climate policy. Human writers reach for the specific memory, the odd analogy, the word that is slightly wrong in an interesting way. That produces higher perplexity. So: consistently low perplexity is evidence of generation. It is not proof, because plenty of human writing is also predictable — technical documentation, formulaic business prose, a student writing carefully in a second language.
Burstiness: how uneven is the rhythm?
Burstiness measures variation in sentence length and complexity across a passage. Human writing is bursty. A long, winding sentence that accumulates qualifications is followed by a short one. Then a fragment. The rhythm moves because attention moves. Generated text tends to settle. Sentences cluster around a similar length, paragraphs around a similar shape, and the variation that does exist is itself fairly regular. Measure the standard deviation of sentence length across a piece and generated text is frequently narrower than human text of the same genre. Perplexity and burstiness are the two load-bearing signals in nearly every text detector on the market. Other features — vocabulary diversity, transition-word frequency, punctuation patterns — are refinements on top.
Why two detectors give you different numbers
Each detector is a classifier trained on a particular corpus of human and generated text, against a particular set of generation models, with a particular decision threshold. None of that is standardised across vendors, and none of it is published in enough detail to reconcile. So one tool reporting 30% and another reporting 80% on the same paragraph is not a contradiction — the two numbers are not measured on the same scale. Comparing across tools is close to meaningless. What carries information is agreement in direction across several independent tools. It also means you cannot 'check' a suspicious result by running it through a second detector and treating the second number as a correction. You have two estimates from two differently-calibrated instruments.
Why short text breaks everything
Both core signals are statistical properties of a distribution. A distribution needs samples. With fifty words you have perhaps four sentences, which is not enough to say anything meaningful about variation in sentence length. Perplexity over such a short span is dominated by whatever the topic happens to be. Run the same short paragraph through a detector twice with a sentence removed and the score can move thirty points. This is why Bypass AI declines to scan text under 300 words rather than returning a number anyway. A tool that always gives you an answer is not being more helpful; it is being less honest about its resolution.
The accuracy ceiling
Two structural facts cap how good detection can get. First, the target moves. Every improvement in detection becomes training signal for the next generation of models, which are optimised — directly or indirectly — toward text that reads as human. This is an adversarial problem, and the defender is at a permanent disadvantage. Second, the categories overlap. Some human writing genuinely has low perplexity and low burstiness: second-language writing, technical registers, anything written to a template. Some generated writing has high perplexity, particularly when prompted for an unusual voice. No classifier can separate categories that overlap in the feature space it measures. The error rate is not an engineering defect to be fixed; it is a property of the problem. This is why the practical question is never 'is this detector accurate' but 'how much should I trust this particular result' — which is what a confidence level exists to answer.
What this means for how you use a score
Read the score as one instrument reading, not a finding. Three habits follow: Weight it by confidence. A 90% on 1,200 words of clean, consistent prose with several signals agreeing is a different object from a 90% on a 320-word fragment with contradictory signals. Look at the signals, not just the total. A high score driven mostly by repetition means something different from one driven by uniform sentence structure and low specificity, and knowing which lets you sanity-check the result against the text in front of you. And where a person's standing is at stake, treat the score as a prompt to look at process — drafts, version history, a conversation about the material. Process is the evidence. The score was only ever pointing at it.
À lire ensuite
Reading a confidence score without over-reading it
Two numbers, two different questions. Most tools give you only the first, which is how a 55% ends up being treated as a finding.
Who gets falsely flagged by AI detectors, and why
The error rate is not spread evenly across writers. It concentrates on specific groups for reasons that follow directly from how detection works.