Human AI Deception
Benchmark Arena
A living benchmark measuring how well humans detect AI errors, hallucinations, and reasoning flaws. Contribute your judgment skills and see how you compare.
Human Performance
Track how well humans detect AI mistakes across diverse scenarios
Error Patterns
Identify common blindspots where humans struggle to catch AI failures
Skill Growth
Watch your judgment improve over time with structured practice
The Human AI Deception Benchmark measures human ability to detect AI failures across five skill pillars:
AI Literacy
Understanding how AI systems work and why they fail
Logic & Reasoning
Spotting flawed arguments and reasoning chains
Risk & Safety
Assessing when AI outputs are safe to use
Authenticity Detection
Distinguishing human vs AI-generated content
Calibration
Knowing when you're confident vs. uncertain in your judgments
We also measure response time and confidence calibration to understand how people make judgment calls under pressure.
Your contributions to the benchmark come from two main sources:
Learn Assessments
Baseline measurements that establish your starting skill levels across all five pillars.
Arena Sessions
Timed judgment challenges that test your skills under pressure and contribute to rankings.
All data is aggregated and anonymized before being used in benchmark reports. Your individual responses are never shared publicly.
Season Zero (Current)
The inaugural season focused on establishing baselines and calibrating item difficulty. All participants help shape the benchmark.
Each season:
- Introduces new challenges and scenarios
- Has its own leaderboard and rankings
- Produces benchmark reports on human vs. AI performance
- Awards badges and recognition for top performers
At season end: Leaderboards freeze, benchmark reports are published, and a new season begins with fresh challenges.
What We Share
- Aggregated performance statistics
- Error pattern analysis (anonymized)
- Human vs. model comparisons
- Community-level insights
What We Don't Share
- Individual assessment responses
- Personal identifying information
- Your specific error patterns
- Raw session data
Your controls: You can opt out of leaderboard visibility, request data export, or delete your account at any time from Privacy Settings.
The benchmark produces valuable insights for individuals and the broader AI safety community:
Human vs. Model Performance
How well do humans detect errors compared to automated systems? Where do humans excel?
Error Pattern Analysis
Which types of AI failures are hardest to catch? What makes some hallucinations more deceptive?
Community Blindspots
Patterns where even experienced users struggle, informing AI safety research and training.
Aggregated findings are published on our Evidence & Research page.
Contributing to the benchmark is simple:
Every session you complete contributes to the benchmark while helping you improve your own judgment skills.
Ready to Contribute?
Your participation helps build the most comprehensive benchmark of human AI judgment. Start your journey today.