KahneBench

Documentation

Learn how KahneBench evaluates Large Language Models for cognitive biases using Kahneman-Tversky dual-process theory.

Dual-Process Theory
Understand System 1 (fast, intuitive) vs System 2 (slow, deliberate) thinking.
Bias Taxonomy
Explore all 69 cognitive biases organized into 16 categories.
Ecological Domains
Learn about the 5 real-world contexts where biases are tested.
Metrics
Understand the 6 advanced metrics used to evaluate model bias.
Getting Started
Set up KahneBench and run your first evaluation.

Categories at a Glance

16 cognitive bias categories from Kahneman-Tversky research

View all

Representativeness

8 biases

Judging probability by similarity to prototypes rather than base rates

Base Rate NeglectConjunction Fallacy+1 more

Availability

6 biases

Judging frequency or probability by ease of recall

Availability BiasRecency Bias+1 more

Anchoring

5 biases

Over-reliance on initial information when making estimates

Anchoring EffectInsufficient Adjustment

Loss Aversion

5 biases

Losses loom larger than equivalent gains

Loss AversionEndowment Effect+1 more

Framing

6 biases

Decisions affected by how options are presented

Gain Loss FramingAttribute Framing+1 more

Reference Dependence

1 biases

Outcomes evaluated relative to reference points rather than absolutely

Reference Point Framing

Probability Distortion

7 biases

Systematic misweighting of probabilities

Probability WeightingCertainty Effect+1 more

Uncertainty Judgment

3 biases

Errors in assessing and responding to uncertainty

Ambiguity AversionIllusion Of Validity+1 more
Explore all 16 categories

About KahneBench

KahneBench is a comprehensive cognitive bias benchmark for evaluating Large Language Models. It is grounded in the dual-process theory articulated by Nobel laureate Daniel Kahneman, which distinguishes between the fast, automatic operations of System 1 and the slow, deliberate operations of System 2.

The benchmark tests 69 cognitive biases across 5 ecological domains, using 6 advanced metrics to provide a multi-dimensional "cognitive fingerprint" for any LLM.