How the detector works, and how it fails.

The category of AI detection has earned a reputation for opaque vendors making precision claims they can't substantiate. We aim to be the opposite of that, this page is the substantiation.

1. What the detector actually does

The detector takes a passage of text and returns a probability that it was generated by a large language model. Internally it combines three signals:

1. Statistical features. Perplexity (under a reference language model), burstiness (sentence-length variance), punctuation density, and n-gram frequency patterns.
2. Learned embeddings. A fine-tuned classifier operating over sentence embeddings from a multilingual encoder.
3. Model-family fingerprinting. Separate classifiers for the major model families (GPT, Claude, Gemini) that identify text patterns specific to each.

The three outputs are combined by a learned gating layer. The gating weights are audited quarterly against the benchmarks below.

2. Benchmark datasets

We publish performance on the following benchmarks. Each is a paired AI/human test set; we do not train on these, they are held out.

• HC3-academic, 3,000 essay-length human/AI pairs across six undergraduate disciplines. Public benchmark.
• GPTZero-bench-2024, 1,200 mixed human-and-AI academic essays. Public benchmark.
• aiessaydetector-internal, 5,000 essay-length pairs we collected through paid-consent data partnerships with community colleges. Licensed, not scraped. Not public.
• multilingual-100, 100 essays per language across five non-English languages, with native-speaker matched AI pairs. Used for stratified FPR reporting.

3. Performance (last updated 2026-Q1)

On the combined benchmark, at the default threshold:

• Accuracy: 94.1%
• Precision (AI class): 93.2%
• Recall (AI class): 91.7%
• F1 (macro-avg): 0.935
• AUC: 0.971
• False positive rate (overall): 1.8%

Stratified false-positive rate

Our headline FPR of 1.8% is the average across all populations in the benchmark. Real-world FPR varies by writer. We report stratified FPR because this is the number that matters to students in known risk groups.

• Native English writers, essays ≥ 400 words: 1.1%
• Non-native English writers, essays ≥ 400 words: 4.7%
• Essays under 300 words (any writer): 6.2%
• Highly formal academic writing (edited ≥ 3 times): 3.4%
• Writers who self-identified as autistic: 3.8%

These numbers are higher than our headline. That is the point of publishing them.

4. What the detector should not be used for

• As the sole evidence in an academic-integrity proceeding.
• On passages under 100 words for emails, or under 300 words for essays. The confidence bands are too wide.
• On non-English text without the language-appropriate classifier enabled.
• On text that's been machine-translated, the translation artifacts interact with detection features.

5. What we do not use training data for

User submissions are never used to train the detector. Training data is licensed (from data partners and public-benchmark datasets) and disjoint from user-submitted text. We will sign contractual prohibitions on training with customer data at the institutional tier.

6. Update cadence

This page is reviewed every quarter. Performance numbers are regenerated against the held-out benchmarks and the stratified FPR is recomputed. Major model releases (GPT-5 class, Claude 5 class, Gemini 3 class) trigger an out-of-cycle update.

7. How to ask for more

Institutional customers: the full methodology pack is shared under NDA during procurement. Researchers: we respond to good-faith research requests; email hello@aiessaydetector.ai with the research question and intended use.