# AI detection accuracy, explained without the marketing

> A single accuracy number hides a lot. What it measures, what it leaves out, and the three questions to ask any detector before you trust its score.

Canonical page: https://winstonai.ai/blog/ai-detection-accuracy-explained

By [Maya Okafor](https://winstonai.ai/authors/maya-okafor), Head of research. Published 2026-10-14. Research, 8 min read.

Every AI detector advertises an accuracy figure, usually somewhere north of 95%. The number is real, but on its own it tells you surprisingly little. Accuracy depends on what was tested, how long the samples were, which models wrote the AI half of the set, and where the line between AI and human was drawn.

This post walks through how we measure our own model, what the headline figure does and doesn't cover, and how to read any vendor's numbers, including ours, with a healthy amount of doubt.

## What accuracy actually counts

In the simplest version, you take a labelled set of documents, half written by people and half generated by language models, and count how many the detector gets right. If it labels 994 of 1,000 documents correctly, that's 99.4% accuracy.

The catch is that one number blends two very different mistakes. Calling an AI essay human is a miss. Calling a human essay AI is a false accusation. A detector can post a great accuracy score while still making the second kind of mistake far too often.

> Accuracy tells you how often a detector is right. The false positive rate tells you how often it's unfair.
>
> Maya Okafor, Head of research

## The three numbers worth asking for

1. The false positive rate: the share of human-written documents flagged as AI. Ours is 0.6% on our current public benchmark.
2. Recall by model: how often it catches text from each major model family, not just the oldest or easiest one.
3. Results on short and edited text: passages under 300 words, and AI drafts a person has revised, are where most detectors struggle.

## Why short text is harder

Detectors look for statistical patterns across many sentences. A 120-word paragraph simply doesn't contain enough signal to be confident either way. That's why our reports show a confidence band, and why we won't return a score at all below a minimum length.

## How to use a score responsibly

A score is evidence, not a verdict. Treat a high AI probability as a reason to look closer: read the highlighted sentences, compare with the writer's earlier work, and ask how the piece was drafted. In our experience, a five-minute conversation resolves more cases than any number.

We publish our full benchmark, including the cases we get wrong, on the research page. If a detector won't show you its misses, that's worth noticing.

## About the author

Maya Okafor, Head of research. Maya leads the team that builds and tests our detection models. Before that she spent eight years in computational linguistics.