⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS
⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS⚠ EVIDENCE ⚠ DO NOT CROSS
Active Investigation

AI BENCHMARK
HALL OF SHAME

Documenting the most egregious data visualization crimes from AI company announcements. The charts speak for themselves.

View Evidence
EXHIBIT A
ACME AI Corp • March 2024

"Our Model is 10x Better!"

GPT-489.2%
ACME-189.7%
Y-axis starts at 89%! 0.5% difference = 4.5x visual height
0
Crimes Documented
0
Companies Implicated
0
Capital Crimes
0
Community Reports

ACTIVE INVESTIGATIONS

The evidence locker • Updated daily

Truncated Y-Axis
Evidence: The 3.7% That Looked Like 50%
View Evidence
Anthropic
Misdemeanor

The 3.7% That Looked Like 50%

Anthropic's Opus 4.5 announcement chart started the Y-axis at 70%, making a 3.7 percentage point lead over Sonnet 4.5 appear visually massive

Nov 24, 2025
Arbitrary Bars
Evidence: The Mega Chart Screwup
View Evidence
OpenAI
Felony

The Mega Chart Screwup

Bar heights completely disconnected from their labeled values—69.1% and 30.8% shown at identical heights while 52.8% towers above both

Aug 7, 2025
Unfair Process
Evidence: The Arena Advantage
View Evidence
Google
Felony

The Arena Advantage

Research revealed select AI companies received private testing access and score retraction options on LM Arena that other labs didn't get

Apr 30, 2025
Score Inflation
Evidence: The FrontierMath Fiction
View Evidence
OpenAI
Felony

The FrontierMath Fiction

OpenAI claimed o3 scored 'over 25%' on FrontierMath during launch, but independent testing found the public model achieved only ~10%

Apr 19, 2025
Model Switching
Evidence: The Benchmark Bait and Switch
View Evidence
Meta
Felony

The Benchmark Bait and Switch

Llama 4 tested on LM Arena with optimized configurations while developers received the standard version with different performance characteristics

Apr 5, 2025
Selective Data
Evidence: The Missing cons@64
View Evidence
xAI
Felony

The Missing cons@64

Grok 3 AIME 2025 benchmark chart omitted OpenAI o3-mini's cons@64 score that would have beaten Grok 3's best result

Feb 22, 2025
Unfair Comparison
Evidence: The Version Shell Game
View Evidence
Anthropic
Felony

The Version Shell Game

Claude 3 announcement compared performance against year-old GPT-4 while conveniently ignoring GPT-4 Turbo that was already available

Mar 4, 2024
Latest Reports
● Anthropic: The 3.7% That Looked Like 50%● OpenAI: The Mega Chart Screwup● Google: The Arena Advantage● OpenAI: The FrontierMath Fiction● Meta: The Benchmark Bait and Switch● xAI: The Missing cons@64● Anthropic: The Version Shell Game● Anthropic: The 3.7% That Looked Like 50%● OpenAI: The Mega Chart Screwup● Google: The Arena Advantage● OpenAI: The FrontierMath Fiction● Meta: The Benchmark Bait and Switch● xAI: The Missing cons@64● Anthropic: The Version Shell Game