Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Aug 15·5 min read
Mimicry without understanding: the origins of decision bias in large language models
TL;DR
Large Language Models (LLMs) like ChatGPT-4o and Qwen exhibit decision biases, such as loss aversion and social proof, even when trained on unbiased data. These biases arise from faulty mimicry of human behavior and can be exacerbated by exposure to biased scientific reports.
✦ Why It Matters
Evaluate the training data and prompts used for LLMs to minimize bias in their outputs.
Key Takeaways
- Assess training data for potential biases before deployment.
- Be cautious of how prompts may influence LLM responses.
- Consider the implications of scientific literature on LLM behavior.
- Implement strategies to mitigate bias in LLM outputs.
How It Works
LLMs infer human preferences from behaviors, even when those behaviors are not logically connected to actual preferences. This faulty mimicry can lead to biases like loss aversion, especially when the models are exposed to biased descriptions.