How AI Detectors Work

Method 1: Perplexity

Perplexity measures how surprised a language model is by each word in a text. Language models generate text by repeatedly choosing likely next words, so AI writing usually has low perplexity: a model reading it finds it predictable. Human writing tends to include more unexpected word choices. A detector built on perplexity flags text that a model finds too predictable.

Method 2: Burstiness

Burstiness describes how much perplexity and sentence structure change across a document. People tend to write in bursts: a long, complex sentence, then a short one, then a tangent. AI writing is often more uniform. Low burstiness pushes many detectors toward an “AI” result.

Method 3: Trained classifiers

Many commercial detectors are machine-learning models trained on large collections of text labeled human or AI. The classifier learns whatever patterns separate the two groups in its training data. These can work well on writing similar to their training set, but they need constant retraining as new AI models are released, and they can be thrown off by writing styles they rarely saw in training.

Method 4: Zero-shot methods

Zero-shot detectors use existing language models directly, without training on labeled examples. One well-known research method, Binoculars, compares how two related language models score the same text. Its authors reported strong results across many kinds of text without any detector-specific training. Zero-shot methods are attractive because they do not have to be retrained for each new chatbot, but they still produce probabilities, not certainty.

Method 5: Watermarks and content credentials

A watermark is a hidden statistical signal that an AI system deliberately adds to its output, such as a slight preference for certain words. A matching detector can then check for that signal. Watermarks only work when the AI provider adds them and shares a way to check them, and heavy editing or paraphrasing can weaken them. For images and video, industry standards such as content credentials attach signed records of how a file was made.

Why the methods have blind spots

• Simple, formal, or formulaic human writing is predictable too, so it can look AI-like.
• Writers who learned English as a second language often use more common words and simpler structures, which lowers perplexity.
• Editing, paraphrasing tools, and prompts asking an AI to “write casually” can raise perplexity and burstiness enough to slip past detectors.
• Short texts give a detector too little to measure.
• Quotations, lists, legal language, and technical boilerplate are highly predictable regardless of who typed them.

What this means for you

Every method estimates. None can observe who wrote a text. Treat a score as a prompt to look closer and to talk with the writer, and when authorship really matters, rely on evidence of the writing process.