AI detection performance
RAIDACL 2024
98.3%
accuracy on RAID
MGTDICAIE 2025
#1
AI detector on MGTD
Accuracy & benchmarks
How accurate is It’s AI for English texts?
Independent benchmarks show how well It’s AI detects AI-written text — and how often it falsely flags authentic human writing.
Human-writing safety
ASAP 2.0Student writing
0.8% FPR
on 25,000 authentic student essays
Third-party verified
It’s AI’s performance is supported by published research and independent third-party benchmarking. Our MGTD results are reported in a peer-reviewed benchmark study, while our RAID result is listed on the official public leaderboard. Both evaluations measure English-language detection.
MGTD · ICAIE 2025
The most accurate AI detector according to MGTD
MGTD aggregates 15 benchmark datasets and nearly 2 million test samples into a direct comparison of leading AI detectors. It’s AI achieved the highest ROC-AUC, ahead of GPTZero, Originality AI, ZeroGPT and other evaluated systems.
- 15 datasets
- ~2M samples
- Multiple detectors compared
RAID · ACL 2024
98.3% accuracy on RAID’s official leaderboard
RAID stress-tests AI detectors across LLMs, writing domains, decoding settings and adversarial modifications — including paraphrasing, misspellings, homoglyphs, whitespace attacks and other realistic perturbations.
- 6M+ generations
- 11 LLMs
- 8 domains
- 11 attacks
English benchmark results
Each benchmark links to the underlying paper, dataset or official benchmark page.
| Benchmark | What it tests | Samples | It’s AI result |
|---|---|---|---|
| MGTD Unified detector comparison | Cross-dataset detector comparison | ~2M samples | #1 Rank0.926 ROC-AUC |
| RAID Official public benchmark | Robust AI detection across models, domains and attacks | 600K+ test samples | 98.3% Accuracy |
| ASAP 2.0 Authentic student essays | Human-writing false positives | ~25K human texts | 0.8% FPR |
Methodology
We report dataset-specific metrics rather than averaging benchmarks that measure different tasks.
Evaluation rules
Language-specific evaluations use the production detector configuration where applicable. Third-party benchmark and leaderboard results follow the benchmark’s published protocol.
Minimum supported text length
Human-writing tests apply It’s AI’s product minimum of 200 characters when filtering is required, so texts the application itself would reject do not distort the reported false positive rate.
Metric definitions
Accuracy is the proportion of correct classifications on datasets containing both human and AI texts. The false positive rate (FPR) measures how often authentic human text is incorrectly flagged as AI. ROC-AUC measures how well the detector distinguishes the two classes across decision thresholds.
