How AI Detection Works

Step 1

The Raw Data

Each dot represents a piece of writing—an essay, a paragraph, a response. Some were written by students. Others were generated by AI.

Right now, they're mixed together with no obvious pattern. To build a detector, GPTZero needs to find a way to separate them—to discover what makes human and AI writing different under the surface.

Step 2

Inside One Sample

Let's zoom into one dot and see what the model actually does with it.

First, the text is split into tokens—small chunks of words and punctuation that the model can process.

The model then processes all the tokens together and outputs a single list of numbers—a compact fingerprint of the writing style. Two texts that are written similarly will produce similar lists.

Step 3

Distributions

When the model runs every sample and converts each one into a list of numbers, a pattern appears: texts with similar writing styles produce similar numbers, so they end up near each other.

Most samples cluster near the center of their group, with fewer toward the edges. This creates a bell-curve shape called a distribution.

The two distributions overlap in the middle—some human writing produces numbers that look like AI output, and vice versa. That overlap is what makes detection hard.

Step 4

The Decision Boundary

The model places a dividing line between the two groups. Everything on one side is classified as human; everything on the other as AI.

Drag the purple line left or right to move the threshold. Watch what happens to the red dots.

Dots that turn red are errors; samples on the wrong side of the boundary.

This is the core tradeoff. GPTZero prioritizes minimizing false positives; it is more important that a student is not wrongly flagged than that every AI text is caught.

Recap

Putting It Together

What you just saw is a simplified version of how AI detection works under the hood:

1. 1Start with a collection of writing samples
2. 2Run each text through the model to get a list of numbers that captures its style
3. 3Similar texts cluster together, forming overlapping bell-curve distributions
4. 4Choose a decision boundary that balances false positives against false negatives

Real detection models are more complex—they use deep neural networks and train on millions of documents—but these are the same core principles at work.

Score 80%+ on the quiz to earn your AI Technical Foundations certificate.