Cognitive Bias in Decision-Making with Large Language Models (BiasBuster)

Authors: Jessica Maria Echterhoff et al. (UC San Diego) Canonical URL: https://arxiv.org/abs/2403.00811


Summary

The most comprehensive single empirical study testing multiple cognitive biases simultaneously in LLM decision-making. Introduces BiasBuster, a framework + dataset of 13,465 prompts designed to probe specific cognitive biases in LLMs, evaluate mitigation strategies, and propose a self-debiasing approach.

This is the closest existing work to the empirical program proposed in the wiki’s hypotheses page.


Biases Tested (All Already Covered in This Wiki)

The paper organizes biases into three categories. The categories that map to wiki concept pages:

BiasBuster categoryWiki concept page
Prompt-induced (e.g. anchoring via primed numbers)Anchoring Bias
Sequential (e.g. order effects)Availability Heuristic / Anchoring
Inherent (e.g. framing-driven response shifts)Framing Effect
Confirmation-style probingConfirmation Bias

Key Findings

  1. Cognitive bias is present across commercial and open-source models. Not just a small-model or unaligned-model phenomenon — present in GPT-4-class systems.
  2. Effect sizes vary by bias. Anchoring and framing show strong effects; some others are subtler.
  3. Self-debiasing works. A novel method asking LLMs to identify and counter their own bias in prompts is effective, without requiring per-bias manually-crafted examples.
  4. Mitigation transfers. A single self-debiasing prompt strategy reduces multiple bias types — a kind of Key Assumptions Check applied to the model’s own reasoning.

Relevance to This Wiki

This paper is the single best empirical reference for the LLM versions of:

It is also partial empirical support for H3 (KAC prevents anchoring propagation). The self-debiasing prompt Echterhoff proposes is functionally a KAC variant — “what assumptions are baked into this prompt?” — applied to bias rather than to evidence.

Open question raised: Echterhoff’s mitigation works on isolated decisions. Whether the same approach holds up in multi-turn agentic chains (where bias propagates across context) is untested.

See Also