The technology · explained plainly
AI detectors don't “know” who wrote something. They estimate how statistically predictable a passage is, using signals like perplexity and burstiness plus machine-learning classifiers. Understanding those signals shows exactly why detectors are useful — and why they get things wrong.
Every AI detector rests on one assumption: text produced by a language model tends to be more statistically predictable than text written freely by a person. Detectors therefore measure how predictable a passage is and translate that into a “likelihood of AI” score. They never actually observe who typed the words — they infer it from patterns, which is precisely why they can be fooled and why they can be wrong.
The signals
Most tools combine some of these. Each is a useful clue and none is proof on its own.
How surprising each word is given the words before it. AI text is often low-perplexity — very predictable — which detectors read as a machine signal. But careful, plain human writing is also predictable, so low perplexity does not prove AI authorship.
How much sentence length and complexity vary across a passage. Humans tend to write in bursts: a long, winding sentence followed by a short one. Uniform, evenly measured rhythm looks machine-made, so low burstiness raises a detector's suspicion.
A model trained on many human and AI samples that outputs a probability for new text. It can catch patterns humans miss, but it is only as good as its training data and can misjudge writing styles, topics or languages it saw little of.
Perplexity, burstiness and classifiers all guess after the fact. Watermarking flips this around. At generation time, a model can nudge its word choices in a subtle, statistical pattern — invisible to a reader but detectable by a matching tool. Google's SynthID-style techniques are the best-known example of this idea for text and media.
Watermarking can be more reliable than guesswork, but it has hard limits:
So watermarking improves detection for some content but does not make detection universal or foolproof.
Classifier-based detectors learn from examples. If a detector was trained mostly on English essays from one kind of writer, it will be least reliable on everything unlike that: technical writing, creative styles, other languages, and — importantly — English written by non-native speakers. Models also drift out of date as new AI systems appear, so a detector trained last year may misjudge this year's output. Training data is both the source of a detector's power and the root of its blind spots.
A false positive is human writing flagged as AI. These are not rare edge cases — they follow directly from measuring predictability instead of authorship:
These false positives are the single strongest reason no serious institution should treat a detector score as sole proof.
Vendors often quote high accuracy figures, but those are typically measured under favourable conditions — long, clean, single-source passages. Real-world accuracy is lower, especially on:
Because both false positives and false negatives occur, a detector is a rough triage tool, not a verdict. Treat any advertised percentage skeptically.
A detector score is an estimate of the probability that text resembles AI-generated writing, based on statistical patterns. A “90% AI” result means the text looks highly predictable to that particular tool — not that a machine definitely wrote it. Run the same text through a different detector and you will often get a different number. Update the tool and the number can change again. That is why a score is a signal to investigate, never proof by itself.
AI detectors are imperfect and produce false results in both directions. A score is not evidence of misconduct, and this site does not endorse using humanizers to hide dishonest work. If you write with AI, keep your work honest and your process transparent — that protects you far better than any number.
Genuine, human-written work does get flagged. If it happens to you:
If you want to compare how different assistants phrase the same idea while you draft, a multi-model workspace like MultipleChat makes side-by-side comparison easy — handy for judging quality rather than chasing a score.
FAQ
Most detectors estimate how statistically predictable a passage is. Two common signals are perplexity, which measures how surprising each word is given the words before it, and burstiness, which measures how much sentence length and complexity vary. Many detectors also feed text into a machine-learning classifier trained on human and AI samples that outputs a probability. None of these signals identify an author with certainty.
Watermarking embeds a subtle, statistical signature into a model's output at generation time, and a matching detector such as Google's SynthID looks for that signature later. It can be more reliable than guesswork, but only for text produced by the specific model that added the watermark, and it can weaken after heavy editing or paraphrasing. It does nothing for text from models that never watermarked their output.
Because they judge predictability, not authorship. Plain, well-structured writing, formulaic formats and non-native English can all look statistically predictable and get flagged as AI even when a human wrote every word. Famous human-written texts have been wrongly flagged in tests. False positives are an inherent limitation of the statistical approach, not a rare glitch.
Accuracy varies widely by tool, text length and text type, and vendor accuracy figures are usually measured under favourable conditions. Real-world performance is lower, especially on short passages, edited or paraphrased text, and writing by non-native speakers. Because both false positives and false negatives occur, no detector should be treated as a reliable single source of truth.
A score is an estimate of the probability that text resembles AI-generated writing based on statistical patterns. A high score means the text looks predictable to that tool, not that it was definitely written by AI. Different detectors give different scores for the same text, and scores change as tools are updated, so a percentage is a rough signal rather than proof.
Stay calm and gather evidence of your process: keep drafts, outlines, version history, and document metadata that show how the work developed. Ask which detector was used and point out that detectors produce false positives, especially on plain or non-native English writing. Request a fair review rather than accepting a single score as proof, and going forward save your revision history routinely.
Explore