How AI detection works, in short
Here is how AI detectors work: they estimate how likely it is that a language model wrote a piece of text. Unlike a plagiarism checker, they usually don't compare your words against a database of sources, and they have no idea who typed them. They measure patterns instead: how predictable each word is, how evenly the sentences run, and how much the text resembles the human and AI examples the detector learned from. What comes out is a probability, not a fact.
Most tools combine several of these signals. You can see the idea on our free AI detector: every sentence is marked as reading human or AI, and the score is the share of the text that reads human. The rest of this guide explains what sits behind a result like that.
What AI detectors look for: perplexity and burstiness
A language model writes one word at a time, and at each step it tends to pick a likely next word. A detector can run a similar model over your text and ask, word by word, how surprised it is. Averaged over the whole text, that surprise is called perplexity.
Take "I drank a cup of ___". "Coffee" or "tea" barely surprises a model. "Thunder" surprises it a lot. Text in which nearly every word is the expected one has low perplexity, and low perplexity is one of the classic hints of machine writing.
The second signal, burstiness, measures how much a text varies from sentence to sentence. People tend to write in bursts: a short sentence, then a long one with a clause or two, then something in between. AI drafts often keep a steady, even rhythm. Compare these two:
- Low burstiness: "The plan improves efficiency. It also reduces costs. Teams can work faster. Customers benefit as well." Four sentences, four words each, all built the same way.
- High burstiness: "The plan saves money. Not much in the first quarter, since the new software has to be paid for up front, but by summer it shows up in every team's budget. Customers notice too."
Detectors also pick up surface habits that models share: the same transitions ("Furthermore", "Additionally"), tidy lists of three, a neat summary line at the end of every paragraph. None of this is proof. A recipe, a legal clause or a lab report is supposed to be predictable, and some people simply write in even sentences. A statistical signal is a hint, which is why many tools add a second method on top.
Trained classifiers: a model that learned from examples
The second approach is a classifier: a model trained on a large set of texts labeled "human" or "AI" until it learns which features separate them. Nobody gives it a rule like "low perplexity means AI". It finds its own patterns, including some no person could put a name to.
OpenAI's own detector, released in January 2023, was built this way: a language model fine-tuned on pairs of human-written and AI-written texts on the same topic. OpenAI withdrew it about six months later, and what happened to it is covered in are AI detectors accurate?
A classifier is only as good as its training data. It does well on text that resembles what it saw in training. On text that doesn't, such as output from a newer AI model, an unusual genre or a language it rarely saw, it gets less reliable. In the same post, OpenAI warned that on inputs very different from its training set, its classifier was sometimes "extremely confident in a wrong prediction".
Document-level vs sentence-level scoring
Some detectors return one verdict for the whole text. Others score each sentence or short passage and build the overall result from those. The two answer different questions: "is this document AI-written?" and "which parts read like AI?"
Turnitin works the second way. The company says it splits a submission into overlapping segments of a few hundred words, gives each sentence a score between 0 and 1, and averages the scores into a prediction of how much of the document is AI-generated.
Sentence-level scoring helps when a text is part human and part AI, because it shows where the signal comes from. It is also noisier. One sentence carries very little evidence, so a single highlighted line means far less than a whole page of them. Our detector scores both the whole text and every sentence, so you see the overall picture and where it comes from.
Watermarking: a signal added while the text is written
Watermarking turns the problem around. Instead of guessing after the fact, the AI model marks its own output as it writes. Google DeepMind's SynthID text watermarking does this by slightly adjusting the probabilities of the words the model picks. The text reads normally. A detector that knows the watermark settings then checks whether the word choices follow the hidden pattern.
Google announced in 2024 that it watermarks text in the Gemini app with SynthID. It described the method in a paper in Nature and released the code so other developers can add it to their own models. According to DeepMind:
- What it handles: light changes such as cropping the text, changing a few words or mild paraphrasing. It works best on longer, open-ended responses like essays or emails.
- What it struggles with: its confidence drops sharply when text is thoroughly rewritten or translated, and it is weaker on answers to factual questions, where there are few word choices to adjust.
One limit applies to every watermark: a detector can only find a mark that was put there. Text from a model that doesn't watermark looks exactly like text with no watermark, so "no watermark found" tells you nothing about whether AI was used.
How do AI detectors work? The three methods compared
| Method | What it measures | Works best on | Main weak spot |
|---|---|---|---|
| Statistical signals (perplexity, burstiness) | How predictable the words are and how evenly the sentences run | Long, unedited AI drafts | Plain or formulaic human writing can look just as predictable |
| Trained classifier | Patterns learned from labeled human and AI examples | Text similar to its training data | Newer AI models, unusual genres and other languages |
| Watermark check | A hidden pattern the AI model added while writing | Longer text from a model that watermarks | Finds nothing in text from models without a watermark, and fades after heavy rewriting or translation |
Watermark checks only help with text from a model that adds a watermark, read by a detector that knows its settings. So most text you paste into a public detector is judged by the first two methods, often combined.
Why two AI detectors give different scores for the same text
Paste one essay into three detectors and you may get three different numbers. That doesn't mean one of them is broken. They measure in different ways:
- Different training data. Each classifier learned from its own collection of human and AI texts, so each has its own blind spots.
- Different thresholds. One tool may call a text AI once it is 50% sure, another only at 90%. The same evidence gets a different label.
- Different units. "70% AI" can mean the tool is 70% sure the whole text is AI, or that 70% of the sentences read like AI. Those are different claims, even though both appear as a percentage.
- Different versions. Detectors are retrained as new AI models appear, so the same tool can score the same text differently a few months apart.
How often each kind of mistake happens is a separate question, and the published tests are summarized in are AI detectors accurate?
What makes AI text harder to detect
Every method above depends on the text keeping the fingerprint the model left on it. Anything that changes the word choices or the rhythm weakens the signal:
- Editing. Each sentence a person rewrites swaps model word choices for human ones. A heavily edited draft really is a mix, and detectors score it as one.
- Short text. A few sentences don't hold enough words for predictability or rhythm to mean much. That's why many detectors set a minimum length. Ours needs 40 words.
- Translation. Moving text into another language replaces nearly every word choice, which scrambles both the statistical signals and any watermark.
- Paraphrasing. Paraphrasing tools and humanizers change exactly what the statistical methods measure: word choice and sentence rhythm.
The same logic runs the other way. Human writing that happens to be plain, predictable and even can look like AI to a detector. If that has happened to you, read why your writing gets flagged as AI.
What AI detectors cannot know
A detector reads text and nothing else. Everything around the text is invisible to it, and that is often what matters most:
- Who wrote it. A score describes the writing, not the writer. It can't tell a student's own plain prose from a model's.
- How the text was made. Typed from scratch, dictated, drafted with AI and rewritten by hand, or tidied with a grammar tool: the detector only sees the final words.
- Which AI tool, if any. Statistical checks and classifiers judge the writing, not the tool. Only a watermark points to a specific model, and only if that model added one.
- Whether AI use was allowed. Some courses and workplaces permit AI for brainstorming or editing. The detector has no idea what the rules were.
That's why a score should start a conversation, not end one. If you teach, our page on AI checkers for teachers covers how to use one fairly.