On August 7, 2025, OpenAI released GPT-5 with a livestream showcasing its benchmark performance. One chart demonstrated the model's supposed superiority at detecting deception. The irony? The chart itself was fundamentally deceptive, with bar heights bearing no relationship to their labeled percentages.
Chart source: OpenAI GPT-5 Launch Livestream via FlowingData
The Violation
- 1
Bar representing 69.1% is identical in height to the bar for 30.8% (should be more than double)
- 2
Bar for 52.8% is significantly taller than the bar for 69.1%
- 3
No consistent mathematical relationship between numbers and visual representation
- 4
Chart title references 'deception detection' while the chart itself deceives viewers
Why This Matters
When the entire point of a benchmark chart is to demonstrate comparative performance, making visual heights arbitrary doesn't just fail—it actively misleads. This chart was specifically about building trust in AI systems' ability to detect deception, making the meta-irony particularly damaging to credibility.
Community Verdict
"The numbers here were accurate but we screwed up the bar charts in the livestream overnight. Very sorry for the mega chart screwup."
"The labels do not remotely match the bar heights. This is a significant visualization error in a major product launch."
"An 'unintentional chart crime' during the most important announcement of the year. The OpenAI marketing team apologized."
The Defense
Imagined company response
"Sam Altman characterized it as an overnight production error, emphasizing that the underlying numbers were accurate even though the visual representation was completely wrong. OpenAI updated their blog post with corrected charts within 48 hours."
Spot a similar crime?
Help us document chart crimes in the wild