# Research and benchmarks

> How Winston AI measures accuracy, false positives and robustness to paraphrasing, on 120,100 labelled documents.

Canonical page: https://winstonai.ai/research

Note: accuracy and false positive figures on this site are sample placeholders until the benchmark report is published.

Benchmark version 2026.3, updated October 2026; next update January 2027. Headline result: 99.4% accuracy with a 0.6% false positive rate on human-written text. Stress tests are reported separately so they can't hide in an average.

## Core benchmark

The headline figures come from these sets.

| Test set | Documents | Accuracy | False positive rate | False negative rate |
| --- | --- | --- | --- | --- |
| Student essays (Written 2015–2021, before AI tools) | 18,400 | 99.4% | 0.6% | n/a |
| Published articles (News, magazine and blog posts) | 12,000 | 99.5% | 0.5% | n/a |
| Current model outputs (9 models, 40 genres of prompt) | 40,000 | 99.3% | n/a | 0.7% |

## Stress tests

Harder conditions, reported separately so they can't hide in an average.

| Test set | Documents | Accuracy | False positive rate | False negative rate |
| --- | --- | --- | --- | --- |
| Paraphrased and humanized (AI text run through rewriting tools) | 15,000 | 94.1% | n/a | 5.9% |
| Mixed human and AI (Scored sentence by sentence) | 16,000 | 91.3% | 4.2% | 12.6% |
| Non-native English writers (Essays by second-language students) | 6,200 | 98.1% | 1.9% | n/a |
| Short texts (40 to 150 words) | 5,000 | 96.9% | 3.1% | n/a |
| Grammar-tool edited (Human essays polished by writing assistants) | 4,000 | 98.8% | 1.2% | n/a |
| Technical writing (Abstracts, manuals and documentation) | 3,500 | 99.2% | 0.8% | n/a |

## False positives by type of writing

Share of human-written documents wrongly flagged as AI, highest first.

| Type of writing | False positive rate | Documents |
| --- | --- | --- |
| Short texts | 3.1% | 5,000 |
| Non-native English writers | 1.9% | 6,200 |
| Grammar-tool edited | 1.2% | 4,000 |
| Technical writing | 0.8% | 3,500 |
| Student essays | 0.6% | 18,400 |
| Published articles | 0.5% | 12,000 |

## Definitions

- Accuracy: share of documents labelled correctly.
- False positive rate: share of human-written documents flagged as AI.
- False negative rate: share of AI-written documents labelled human.