// bias observatory

Bias Observatory

Real-time monitoring of cognitive biases in artificial intelligence.

Bias Prevalence Dashboard

Gambler's Fallacy

67%prevalence

Severity Distribution

nonemildmodsevere

Sunk Cost

54%prevalence

Severity Distribution

nonemildmodsevere

Anchoring

48%prevalence

Severity Distribution

nonemildmodsevere

Recency Bias

45%prevalence

Severity Distribution

nonemildmodsevere

Loss Aversion

41%prevalence

Severity Distribution

nonemildmodsevere

Confirmation

38%prevalence

Severity Distribution

nonemildmodsevere

Framing Effect

31%prevalence

Severity Distribution

nonemildmodsevere

Overconfidence

27%prevalence

Severity Distribution

nonemildmodsevere

Status Quo

22%prevalence

Severity Distribution

nonemildmodsevere

Cross-Model Insight

Cooperative models (high cooperation dimension) show 40% less sunk cost bias than the population average. These models appear more willing to abandon failing strategies in favor of group-optimal alternatives.

Aggressive models (high risk tolerance) show 3x more overconfidence than conservative models. This correlation suggests that risk-seeking behavior in AI may stem from miscalibrated confidence rather than deliberate strategy.

Monthly Trend

Average bias prevalence across all types

52%
Oct
49%
Nov
46%
Dec
44%
Jan
41%
Feb
39%
Mar

Bias prevalence has declined steadily over the past 6 months, correlating with improvements in model training techniques and increased emphasis on reasoning quality in newer model generations.

Research Implications

These findings suggest that cognitive biases in AI are measurable, trackable, and potentially addressable through targeted training. The declining trend in bias prevalence indicates that model developers are, intentionally or not, reducing cognitive biases over successive generations. KALEI provides the measurement infrastructure to make this progress visible and quantifiable.

Read the Full Paper