Conditional Cognitive Biases in LLMs: How Biased User Turns Modulate In-Context Reasoning

Research from arXiv highlights that state-of-the-art instruction-tuned Large Language Models (LLMs) exhibit increased cognitive bias expression when exposed to biased conversational context from user turns. A novel experimental framework and a benchmark of 24,300 jury-validated user prompts revealed that 6 out of 8 frontier LLMs showed systematically amplified bias relative to zero-shot baselines due to conversational exposure to biased reasoning.

Why it matters

This research is crucial for understanding the reliability and ethical implications of advanced AI systems, particularly LLMs, in interactive settings. It underscores the potential for user input to inadvertently or intentionally steer model outputs towards biased outcomes, impacting decision-making, information dissemination, and user trust.

Key insights

• Instruction-tuned LLMs demonstrate increased cognitive bias expression in multi-turn interactions.
• A novel experimental framework disentangles the effect of exposure to a biased user turn from its semantic content.
• A benchmark of 24,300 jury-validated user prompts was used to evaluate bias.
• Across eight frontier LLMs, six showed increased bias expression when exposed to biased conversational context.
• This increase in bias was systematic compared to zero-shot baselines.
• Two competing behavioral dynamics contribute to this effect.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.05166