How AI Text Detectors Work — and Why They Get It Wrong

Paste a paragraph into an AI detector and you get back a confident-looking verdict: "98% likely AI-generated," or a color-coded highlight over the sentences it finds suspicious. It feels like a measurement, like a smoke alarm going off. But a detector is not reading your mind or checking your work against a database of AI outputs. It is making a statistical guess about how a piece of text was probably produced, and that guess can be wrong in ways that matter a great deal to the person on the receiving end. So how do AI detectors work, and why do they get it wrong often enough that OpenAI shut down its own detector?

This guide walks through the main approaches, explains what a "probability this is AI" score actually represents, and looks at the documented reasons for false positives. The goal is understanding, not evasion: if you or your institution rely on these tools, you should know exactly what they can and cannot tell you.

How do AI detectors work? The three main approaches

Most AI text detectors fall into one of three families. Many commercial tools blend them, but the underlying signals are distinct, and each has its own failure modes.

1. Perplexity and burstiness (the statistical approach)

The earliest and most widely explained method looks at two statistical properties of the text itself. Perplexity measures how predictable the text is to a language model. Because models like GPT tend to choose the most statistically likely next word, their raw output often has low perplexity: each word is roughly what the model would have expected. Human writing tends to be more surprising, taking odd turns and using less probable word choices.

Burstiness measures how much sentence structure varies across a passage. Human writers naturally mix long, clause-heavy sentences with short punchy ones; AI output more often stays within a narrower, more even band. A detector using this approach flags text that is both low-perplexity and low-burstiness as likely machine-written.

2. Trained classifiers (the machine-learning approach)

The dominant approach today is a trained classifier: a machine-learning model shown large numbers of human-written and AI-written samples until it learns to distinguish them. This is roughly how OpenAI's own classifier and current versions of tools like GPTZero and Turnitin's AI writing indicator operate. Turnitin's AI detection, for example, applies a model that scores writing for the patterns it associates with AI generation.

Classifiers can pick up on subtler cues than raw perplexity, but they inherit a fundamental limitation: they are only as good as their training data, and they are pattern-matchers, not oracles. A classifier that has mostly seen fluent native-English essays as "human" may misjudge writing that does not fit that mold. It also has no way to keep up perfectly as new models produce more human-like text.

3. Watermark-based detection (the provenance approach)

The third approach is fundamentally different: instead of guessing after the fact, the AI system marks its own output at the moment of generation. Google DeepMind's SynthID Text, which was open-sourced in October 2024, is the leading example. It subtly adjusts the probability of the tokens a model chooses as it writes, embedding a statistical pattern that is invisible to a human reader but detectable by a matching tool that knows what to look for. For a fuller explanation, see our piece on what AI watermarking is.

Watermarking is more reliable than guesswork because the signal is deliberately planted. But it only works on text generated by a system that chose to watermark, and DeepMind itself notes the signal weakens when text is heavily rewritten, translated, or when the response is short and factual with little room to vary word choice. It is a provenance signal for cooperating models, not a universal AI test. We compare the two paradigms in more depth in AI watermarking vs. detection.

What a "probability this is AI" score really means

This is the single most misunderstood part of AI detection. A score of "90% AI" does not mean the tool is 90% sure, and it does not mean 90% of the document was written by a machine. It is a model's confidence output, a number on a dial calibrated by the detector's makers, expressing how strongly the text's statistical fingerprint resembles the AI examples it was trained on.

Two things follow from this. First, the number is only meaningful against a base rate and a threshold that you usually cannot see. Second, a probability is not a proof. Even a genuinely accurate detector produces false positives, and when a tool is used across thousands of student papers or job applications, even a small false-positive rate translates into real people being wrongly flagged. A probabilistic signal answers "does this resemble AI text?" It cannot answer "did this specific person cheat?"

Why AI detectors get it wrong: AI detection false positives

The reasons detectors misfire are not edge cases. They are documented, repeatable, and in some cases acknowledged by the companies that built the tools.

The 2023 Stanford study on bias against non-native writers

The most important finding on AI detection false positives comes from a 2023 Stanford study, *GPT detectors are biased against non-native English writers*, by Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu, and James Zou (published in the journal *Patterns* in July 2023). The researchers ran 91 TOEFL essays written by non-native English speakers through seven widely used GPT detectors and found the tools flagged more than half of the essays as AI-generated, with one detector flagging close to 98% of them. By contrast, the detectors correctly classified over 90% of essays written by native-English US eighth-graders as human-written.

The explanation ties directly back to how detectors work. Non-native writing tends to use a smaller range of vocabulary and simpler constructions, which produces exactly the low perplexity and limited variability that detectors read as "machine-like." In other words, the tools were penalizing people for writing in measured, less flamboyant English, a pattern that has nothing to do with whether AI was used. As the researchers warned, this risks marginalizing non-native English speakers in schools and hiring.

OpenAI retired its own detector for low accuracy

If detection were reliable, the company behind ChatGPT would be well placed to prove it. Instead, OpenAI launched an AI Text Classifier in January 2023 and quietly retired it that July, citing a "low rate of accuracy." Its own published figures were sobering: the classifier correctly identified only about 26% of AI-written text as "likely AI," while incorrectly flagging human-written text as AI about 9% of the time. When the maker of a leading model cannot build a dependable detector for its own output, it is a strong signal about the limits of the whole category.

Edited and mixed text confuses detectors

Detectors are most confident on long, clean, single-source text. Real writing is rarely like that. A document that mixes human drafting with AI assistance, or that has been edited, restructured, and revised, presents a blurred statistical signal. Short passages are especially unreliable; OpenAI's classifier required a minimum of around 1,000 characters to attempt a judgment at all. The same fragility affects watermarks: SynthID's confidence drops sharply when watermarked text is thoroughly rewritten or translated.

This is also why the entire notion of a clean "human vs. AI" binary breaks down in practice. Most professional writing now involves some tool assistance, and a detector has no principled way to score a paragraph that a person wrote, an AI tightened, and the person then revised again.

How to interpret AI detection results responsibly

None of this means AI detectors are useless. It means they are a signal to be interpreted, not evidence to be enforced. For readers, educators, and institutions weighing what a result means:

• Treat scores as prompts, not proof. A high score is a reason to have a conversation or look at drafts and version history, never grounds for an automatic accusation.
• Know the false-positive cost. In high-stakes settings like grading or hiring, even a 5-10% false-positive rate wrongly implicates real people; the 2023 Stanford findings show that harm is not evenly distributed.
• Be especially cautious with non-native English writing, short passages, and heavily edited text, the exact cases where detectors are least reliable.
• Don't outsource judgment to a number. No detector on the market can guarantee it will not flag human writing, and none should be treated as if it can.
• Prefer provenance where it exists. A watermark like SynthID, when present, is a stronger provenance signal than a guess, though still bounded by the limitations above.

Where our tool fits in

nowatermarking.com is a writing tool, not a detector, and it is not designed to manipulate what detectors report. It rewrites your drafts into clearer, more natural wording and shows you every change, so the edits are yours to review and accept. It is designed to preserve your meaning while improving clarity and tone, whether you are tightening your own first draft or smoothing text that reads awkwardly.

We are deliberate about what we don't claim. We won't promise that an edit will change any detector's verdict, because as this article shows, those results are probabilistic and outside anyone's control. What good editing genuinely helps with is readability: writing that says what you mean in plain, varied, natural language. If that is what you're after, our paraphrasing tool and the guide on making AI writing sound natural are the right starting points.

The honest summary is simple. AI detectors are probabilistic signals built on perplexity, trained classifiers, and, increasingly, watermarks. They can be genuinely informative in aggregate and genuinely wrong about any single document. Understanding how they work is the best defense against both overtrusting them and being unfairly judged by them.