Our Numbers. Your Confidence.
We publish our accuracy data because transparency builds trust. Every number here comes from an independent benchmark of 211 real-world text samples.
Key Metrics
Per-Engine Breakdown
Each engine has different strengths. Consensus combines them.
← Swipe to see all columns →
| Engine | False Positive Rate | True Positive Rate | Strength |
|---|---|---|---|
| GPTZero | 0.0% | 88.2% | Human Guardian — lowest FPR |
| Winston AI | 3.5% | 90.2% | Balanced detector |
| ZeroGPT | 18.4% | 94.1% | Aggressive — highest TPR |
| OmniScore (Consensus) | 2.5% | 96.1% | Best of both: low FPR + high TPR |
FPR = human text incorrectly flagged as AI (lower is better). TPR = AI text correctly identified (higher is better). These figures were measured in February 2026 on engine lineup v1 (GPTZero + Winston AI + ZeroGPT). We retired GPTZero on 14 June 2026 and replaced it with Sapling AI; the current lineup is Winston AI + Sapling + ZeroGPT. We publish the v1 numbers unchanged rather than restate them for a lineup we have not re-measured — the v2.5 benchmark rerun is in progress and will be published here.
Why Consensus Beats Any Single Engine
ZeroGPT alone flags 18.4% of human writing as AI. Through consensus, that drops to 2.5%.
ZeroGPT individual FPR 18.4% → Consensus FPR 2.5%
All three engines agree on the verdict (3/3)
Two of three engines agree, outlier ignored (2/3)
All engines disagree — flagged as uncertain
When engines disagree, that's information too. A split verdict tells you the text is ambiguous — more honest than a false confidence score from a single detector.
How We Built the Benchmark
Rigor in, trust out.
211 Real-World Samples
Human texts from 15+ sources: classic literature, academic papers, student essays, news articles, blog posts, forum discussions, and professional writing. AI texts from 6+ models: GPT-4o, Claude 3.5, Gemini, Llama, Mistral, and more.
Zero LLM Contamination
All human samples collected using pure extraction tools (Firefox Reader Mode, Firecrawl). No LLM was used to 'clean' or 'extract' human text — because LLM extraction produces AI-like artifacts that corrupt benchmark integrity.
Contamination Audit
10 samples initially labeled 'human' were reclassified after all three engines unanimously flagged them — traced to using an LLM for text extraction. Lessons learned, methodology improved, and these samples excluded from scoring.
Bilingual Coverage
171 English + 40 German samples. Both languages tested against all three engines to verify cross-language accuracy.
Continuous Re-evaluation
Every engine upgrade, algorithm change, or threshold adjustment triggers a full benchmark re-run. The dataset grows with every iteration.
Most Accurate AI Detector Tools in 2026
Which AI detector is the most accurate? Individual tools achieve 85-95% accuracy, but they frequently disagree — our benchmark shows engines contradict each other on 15-30% of texts. A single score cannot give you certainty.
OmniDetect solves this with multi-engine consensus. By combining Winston AI (the accuracy anchor), Sapling (token-level, enterprise-grade) and ZeroGPT (the volume workhorse), a single engine's blind spot never becomes your verdict on its own.
| Tool | Engines | FPR | Approach |
|---|---|---|---|
| OmniDetect | 3 (consensus) | 2.5% | Multi-engine verdict |
| GPTZero | 1 | ~9% | Perplexity-based |
| ZeroGPT | 1 | ~8% | Deep learning |
| Winston AI | 1 | ~12% | Transformer-based |
The methodology is simple: when three independent engines agree, the result is far more reliable than any single opinion. It's the difference between one judge and a jury.
Honest Limitations
No AI detector is perfect. Here's what ours struggles with.
Claude mimicry is hard to catch
Two AI samples imitating student and narrative styles scored under 16%. Winston AI and ZeroGPT missed them entirely — only GPTZero flagged them.
Academic writing gets higher scores
All three false positives were academic or professional texts. Formal, structured writing can resemble AI output patterns.
Short texts are less reliable
Texts under 300 words produce less stable results across all engines. We recommend 500+ words for reliable verdicts.
Paraphrasing tools reduce accuracy
Heavily paraphrased AI text may bypass all three engines. No detector in the market fully solves this challenge.
ESL writers may see elevated scores
Non-native English writers sometimes produce patterns that overlap with AI-generated content, leading to higher-than-expected scores.
Frequently Asked Questions
Related Detection Tools
See for Yourself
Numbers are nice. Experience is better. Try a free scan and judge the accuracy firsthand.
Start Free Scan