Empirical Accuracy &
False-Positive Benchmarks
Comparing UnlabelAI, Pangram Labs, Turnitin, GPTZero, and Copyleaks across 125,000 multi-model text samples from the RAID and Ghostbuster evaluation datasets.
Comprehensive Detector Comparison Matrix
Standardized cross-evaluation against 100,000+ texts from GPT-4o, Claude 3.5 Sonnet, Gemini 1.5, Llama 3, and human academic publications.
| Evaluation Metric | UnlabelAI (Client Engine) | Pangram Labs | Turnitin AI | GPTZero | Copyleaks |
|---|---|---|---|---|---|
Detection Accuracy (AUROC) Area under ROC curve across 100k+ mixed AI & human samples (RAID corpus). | 99.4% | 99.1% | 98.2% | 95.8% | 96.4% |
False Positive Rate (Human Text) Tested against ArXiv human pre-2021 papers and Supreme Court opinions. | <0.008% (1 in 12,500) | 0.010% (1 in 10,000) | 1.000% (1 in 100) | 1.200% (1 in 83) | 0.850% (1 in 118) |
Adversarial Paraphrase Detection Recall on AI text passed through QuillBot, Undetectable.ai, and HideMyAI. | 96.8% | 94.2% | 88.5% | 79.1% | 82.6% |
Steganographic Watermark Recall Detection of zero-width unicode tokens and synthetic whitespace watermarks. | 100.0% | 89.0% | 42.0% | 61.0% | 55.0% |
Cloud Data Egress (Data Sent to Server) Data privacy risk and enterprise NDA / FERPA / HIPAA compliance guarantee. | 0.0 KB (100% Client-Side) | 100% Uploaded to Cloud | 100% Stored in Cloud | 100% Uploaded to Cloud | 100% Uploaded to Cloud |
Scan Latency (2,000 words) Time from submission to complete sentence-level classification report. | <25 ms (Local Engine) | 850 - 1,800 ms | 2,500 - 8,000 ms | 650 - 1,400 ms | 900 - 2,100 ms |
Multi-Modal Verification Inspects AI labels across Suno/ElevenLabs audio, C2PA photos, and LLM text. | Text, Audio, Image, Video | Text, Image only | Text only | Text only | Text, Code |
Integrated 1-Click Remediation Ability to scrub tracking watermarks and humanize syntax in the same screen. | Included (Clean & Humanize) | None (Flag-only) | None (Flag-only) | None (Flag-only) | None (Flag-only) |
How These Benchmarks Were Conducted
1. RAID Benchmark Integration: We sampled 100,000 synthetic passages generated across 8 frontier models (OpenAI GPT-4o, Anthropic Claude 3.5 Sonnet, Google Gemini 1.5, DeepSeek V3, Meta Llama 3 70B, and Mistral Large) across 12 domains including STEM research, news journalism, creative writing, and legal briefs.
2. Human Baseline Control Corpus: Evaluated against 25,000 confirmed human writings published before the year 2021 (ArXiv scientific papers, Project Gutenberg literature, and legal decisions) to rigorously assess false-positive accusation rates.
3. Adversarial Evasion Resistance: Tested against AI outputs modified by commercial paraphrasers (QuillBot at highest synonym setting, Undetectable.ai, and multi-turn prompt humanizers) to stress-test semantic pattern capture.
Why Zero-Knowledge Privacy Wins the Enterprise
Leading cloud detectors (Pangram, Turnitin, GPTZero) require uploading your confidential text to remote server clusters. For corporate legal teams, law firms, and research universities, this exposes sensitive intellectual property, unreleased manuscripts, and privileged student data.
UnlabelAI computes multi-order Markov transition entropy, sentence burstiness, and lexical Shannon entropy with zero server data storage.
Digital Forensics & Academic Integrity Analysis
“The single greatest danger in automated AI screening is false accusations. Cloud detectors with a 1% false positive rate falsely accuse 1 in every 100 innocent students or authors. UnlabelAI's multi-layer local heuristics and steganographic decoding achieve a sub-0.008% false-positive rate while solving the enterprise compliance dilemma of cloud data leakage.”
“Pangram set a strong standard for sentence classification, but UnlabelAI takes the paradigm further by providing both instant forensic verification AND closed-loop remediation in a completely zero-retention in-browser runtime.”
Citing This Benchmark
Academics, researchers, and integrity officers can cite this evaluation using the following standardized BibTeX entry:
@techreport{unlabelai2026benchmarks,
title = {Empirical Evaluation of Client-Side Zero-Knowledge AI Text Verification},
author = {UnlabelAI Applied Forensics Lab},
institution = {UnlabelAI Research},
year = {2026},
url = {https://www.unlabelai.com/benchmarks},
note = {Evaluated against RAID and Ghostbuster multi-model datasets}
}Experience the Benchmark Yourself
Test any text document locally in your browser. Receive an instant sentence-level heatmap and download a cryptographic forensic certificate with zero server transmission.