How Do AI Detectors Work? The Honest Explanation
Published July 25, 2026 · FiftyGPT Editorial Team · 8 min read · Updated October 11, 2026
An AI detector reads a piece of writing and estimates how likely it was produced by an AI model such as ChatGPT, Gemini, or Claude rather than a person. You paste text in, and within seconds you get a probability score — usually a percentage — plus clues about which sentences look machine-generated.
What is an AI detector, exactly?
The deeper question — how do AI detectors work, and how much should you trust one? — is what this guide answers. The short version: detectors measure statistical patterns like perplexity and burstiness, they are useful screening tools, and they are far from infallible. See it in action with the free FiftyGPT AI detector — no sign-up, up to 100 checks a day.
At its simplest, an AI detector is a classifier. It has been trained on large collections of two kinds of text: writing known to be human-authored, and writing known to be produced by AI language models. From those examples it learns the statistical fingerprints that tend to separate the two. When you submit new text, it compares your writing against everything it learned and outputs a judgment about which group your writing most resembles.
That judgment is a probability, not a verdict. A result like "85% likely AI" means the patterns in your text closely match the patterns the detector associates with machine writing. It does not mean a machine confessed, and it is not proof. Treating a percentage as a fact is the most common mistake people make with these tools — and the research says it's a costly one: in the broadest head-to-head evaluation available, every one of 14 tested detectors scored below 80% overall accuracy (Weber-Wulff et al., 2023).
One more distinction up front: an AI detector is not a plagiarism checker. A plagiarism checker compares your words against published sources to find copied passages; AI detection studies the writing style itself. The two answer different questions, and confusing them leads to bad decisions. Our guide on the difference between an AI detector and a plagiarism checker lays out when you need which — and FiftyGPT offers both as free tools, since real workflows need both.
What do perplexity and burstiness actually measure?
Most modern detectors lean on two core statistical ideas. Understanding them removes the mystery — once you know what the tool measures, its scores stop feeling like magic.
Perplexity measures how predictable a piece of text is to a language model. AI models are built to choose the most probable next word again and again, which produces writing that is smooth, fluent, and statistically unsurprising. Human writing tends to be less predictable: we reach for an unexpected word, break our own rhythm, and phrase things in ways a model would consider unlikely. Low perplexity pushes a detector toward an AI judgment; higher, more varied perplexity looks more human.
Burstiness measures variation in sentence structure and length. People write in bursts — a long, winding sentence followed by a short one, then a fragment for emphasis. AI-generated text often moves along in sentences of similar length and shape, which reads as unnaturally even. Uniform, low-variation rhythm registers as a machine signal; the natural unevenness of human writing drops the AI probability.
Together, they give the detector a sense of whether the text surprises the way human writing does. Neither is a perfect tell. A heavily edited AI draft can gain burstiness; a naturally plain human writer can score low on perplexity. Combined across a whole document, though, they are a surprisingly useful signal — and they sit at the heart of modern AI detection. For the full research picture on how well that signal holds up, see our AI detection accuracy research roundup. Short passages are a special weak spot, as our guide to detecting AI in short text explains — and the AI detector FAQ answers the follow-up questions readers ask most.
See the signals yourself: paste your own writing into the free FiftyGPT AI detector and check which sentences it highlights — you'll start recognizing the patterns instantly.
What other signals do detectors weigh?
Beyond perplexity and burstiness, detectors weigh a range of smaller stylistic tells. Repetitive sentence openers, an over-reliance on transition words like "moreover" and "furthermore," perfectly balanced paragraph lengths, and a smooth but generic tone all nudge the score upward. So does the absence of the small imperfections that mark real writing: personal anecdotes, specific concrete details, mild inconsistency, the occasional idiosyncratic phrase.
Good detectors also work at the sentence level, not just the document level. Rather than handing you one number, a strong tool highlights which specific passages look most machine-like. That sentence-by-sentence view is far more useful than a single percentage, because it shows you exactly where to focus if you want to revise. The FiftyGPT AI detector provides these sentence-level clues, so you can see the reasoning behind the score instead of trusting a black box.
It is worth knowing what these signals cannot see. A detector has no access to your writing history, your browser, or any hidden metadata. It only reads the words on the page. That is why the same text can score differently across tools, and why short samples are so unreliable — there simply isn't enough signal in thirty words to reach a confident conclusion. (It's also why free detectors and paid ones share the same ceiling: our comparison of the best free AI detectors shows the differences are workflow and cost, not certainty.) Even OpenAI warned that its own classifier was "very unreliable" on text under 1,000 characters before retiring the tool entirely in July 2023 after it correctly identified only 26% of AI-written text in testing (Search Engine Land).
One more signal worth understanding: editing erases the fingerprints. In the Weber-Wulff evaluation, average detection accuracy was 74% on unmodified AI text, but fell to 39% on manually edited AI text and just 22% on machine-paraphrased text. A separate NeurIPS 2023 study found automated paraphrasing collapsed DetectGPT's accuracy from 70.3% to 4.6% (Krishna et al., 2023, via this review). This is the mechanism behind every "humanizer" tool — including the free AI humanizer — and it's why a clean score never proves human authorship.
How accurate are AI detectors in practice?
This is the question that matters most, and the honest answer is: useful, but far from infallible. Here's what the research actually shows, study by study:
- Weber-Wulff et al. (2023) — 14 tools, all below 80% accuracy. Human-written text was correctly identified 96% of the time, but unmodified AI text only 74% — and edited or paraphrased AI text mostly slipped through.
- Liang et al., Stanford (2023) — seven popular detectors misclassified an average of 61.3% of TOEFL essays by non-native English writers as AI-generated, while performing near-perfectly on native-speaker essays (Patterns). The contrast is stark: 61.3% false positives on non-native writing versus 5.1% on native writing — and 19.8% of the human-written TOEFL essays were flagged by all seven detectors unanimously.
- OpenAI's own classifier (2023) — correctly identified just 26% of AI-written text and mislabeled 9% of human text, then was retired for low accuracy after six months.
- Perkins et al. (2024) — in a real university assessment with 22 inserted GPT-4 submissions, Turnitin flagged 91% as containing some AI but marked only 54.8% of the generated content; faculty escalated 54.5% to misconduct procedures.
- Hadra, Cambridge & Mesbah (2026) — in a 192-text benchmark, Turnitin reached 61% overall and Originality.ai 69%, with both struggling badly on mixed human-and-AI writing.
The pattern across all of it: detectors are decent screening instruments on long, unedited, clearly-AI or clearly-human text — and unreliable everywhere else. As the Weber-Wulff team concluded, detection tools "do fail, they are neither accurate nor reliable." For a plain-English verdict on the numbers, see are AI detectors accurate, and for the human cost of the errors, our false positives guide.
When should you trust a score — and when shouldn't you?
Trust a score most when the text is long, unedited, and the result is extreme and consistent at the sentence level — many passages flagged, not one. Treat it as a prompt to look closer: check drafts, ask questions, apply judgment. That's the legitimate use case, whether you're a teacher reviewing essays (our teacher's guide), an editor checking submissions, or a writer sanity-checking your own draft.
Don't trust a score when the text is short, when it's been paraphrased or heavily edited, when the writer is a non-native English speaker or writes in a plain formal style — the groups most exposed to false positives — or when you're using one run from one tool as the entire basis for a consequential decision. No detector on the market, free or paid, can promise certainty — and reputable ones don't.
The practical workflow that respects these limits: run the check on a long sample, read the sentence-level clues, corroborate with what you know about the writing process, and decide like a human — because you are one. If a draft scores high and you want to revise it into more natural writing, the free AI humanizer and paraphraser in the same toolkit handle that. Everything lives in one free place.
Test the theory on your own text: try the free FiftyGPT AI detector — free, no sign-up, up to 15,000 characters per request, with sentence-level clues behind every score.