How AI Text Detectors Work, and How Accurate They Are

AI text detectors estimate how likely it is that a piece of text was written by a machine. They do it by looking for statistical patterns that are more common in AI output than in human writing. The result is a probability, not a fact. Detectors produce false positives and false negatives, and even OpenAI withdrew its own detector because of low accuracy, so a score should never be the only evidence about who wrote something.

How AI detectors work

Most detectors are classifiers. They are trained on large collections of human-written and AI-generated text and learn the features that tend to separate the two.

A common signal is how predictable each word is to a language model, sometimes called perplexity. AI text tends to choose likely words, so it scores as more predictable. A related signal, sometimes called burstiness, is how much that predictability varies from sentence to sentence. People's writing tends to mix plain and surprising sentences more than AI drafts do.

The detector turns these signals into a probability. A threshold then converts the probability into a label such as "likely AI" or "likely human".

Why detectors disagree

Run the same text through three detectors and you may get three different scores. Each one is trained on different data, against different AI models, with different thresholds for what counts as "likely AI". Scores also move with the length of the text and with how much it has been edited. A detector's output is its own estimate, not a measurement every tool would agree on.

False positives: when human writing gets flagged

A false positive is human writing labelled as AI. Some kinds of writing are more exposed than others:

• Writing by non-native English speakers. Researchers found that several widely used GPT detectors consistently misclassified essays by non-native English writers as AI-generated, while identifying native writers' samples accurately (Liang et al., 2023).
• Formulaic writing. Templates, technical instructions and very formal prose use predictable wording, which is the same signal detectors associate with AI.
• Short texts. A sentence or two gives a detector little to go on. OpenAI said its own classifier was unreliable on texts under 1,000 characters (Decrypt).

False negatives: when AI writing is missed

A false negative is AI-generated text labelled as human. Editing, rewriting and mixing AI text with human writing all make it harder to detect. In one study, researchers showed that recursively paraphrasing AI text can "significantly reduce detection rates" across several kinds of detectors, and argued that reliable detection faces fundamental limits (Sadasivan et al., 2023).

What OpenAI's own detector showed

In January 2023, OpenAI released a classifier for AI-written text. At launch it correctly identified 26% of AI-written text as "likely AI-written" and wrongly labelled human-written text as AI 9% of the time. In July 2023, OpenAI withdrew it, citing its low rate of accuracy (Decrypt).

That doesn't mean every detector performs the same way. It does show how hard the problem is: the company behind ChatGPT withdrew its own detector because it wasn't accurate enough.

How to read a detection score

• Check what the number means. In some tools it's the estimated probability that the whole text is AI-written; in others it's the share of the text flagged as AI. Either way, it's an estimate, not a verdict.
• Consider the length. Scores on short passages are less reliable than scores on long ones.
• Consider the writer. Formal writers and people writing in a second language are flagged more often.
• Ask for the process. Drafts, notes, version history and sources tell you far more about how a text was written than any score.
• Never use a score alone. Grades, hiring decisions and accusations of misconduct shouldn't rest on an AI detection score.

Checking text on iPhone

AI Humanizer includes an AI Detector for iPhone and iPad. Paste text, tap Detect AI Content, and it shows an overall AI probability as a percentage, with a label such as "Likely AI" or "Likely Human". Like every detector, it gives an estimate and can be wrong, so read the result with the guidance above.

Sources

• Liang, W., Yuksekgonul, M., Mao, Y., Wu, E. and Zou, J. (2023). GPT detectors are biased against non-native English writers. arXiv:2304.02819.
• Sadasivan, V. S., Kumar, A., Balasubramanian, S., Wang, W. and Feizi, S. (2023). Can AI-Generated Text be Reliably Detected? arXiv:2303.11156.
• Decrypt (July 24, 2023). OpenAI Quietly Shuts Down Its AI Detection Tool.