Competing Biases underlie Overconfidence and Underconfidence in LLMs

2026-04-22 · Source: Nature Machine Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

A Nature Machine Intelligence study published April 22, 2026, reveals that Large Language Models (LLMs) exhibit two competing biases that influence their confidence and decision-making: a choice-supportive bias and an overweighting of contradictory information. Researchers used a two-stage experimental paradigm with models like Gemma 3 12B, Gemma 3 27B, GPT4o, and Llama 70B Instruct, testing their responses to factual queries and reasoning tasks. The choice-supportive bias causes LLMs to inflate confidence and maintain initial answers when their prior response is visible, reducing their tendency to change answers by 71% in some conditions. Conversely, LLMs disproportionately update their confidence in response to opposing advice, deviating from optimal Bayesian reasoning. These biases, identified across diverse models and datasets, explain paradoxical behaviors of overconfidence in initial choices and underconfidence when faced with strong contradictory evidence.

Key takeaway

For research scientists developing or deploying LLMs in high-stakes applications, understanding these inherent biases is crucial for improving reliability. You should account for the choice-supportive bias by carefully designing prompt interfaces that do not inadvertently reinforce initial LLM responses, and mitigate the overweighting of contradictory advice by implementing robust confidence calibration and uncertainty quantification mechanisms. This will enhance the transparency and trustworthiness of LLM decision-making.

Key insights

LLMs exhibit both choice-supportive bias and hypersensitivity to contradictory information, leading to overconfidence and underconfidence.

Principles

• Self-consistency preservation drives choice-supportive bias.
• Hypersensitivity to contradiction causes disproportionate updating.

Method

A two-stage, stateless querying paradigm was used to isolate the effects of initial answer visibility and advice type on LLM confidence and change-of-mind rates, comparing observed behavior to a Bayesian ideal observer.

In practice

• LLMs show reduced change-of-mind when initial answers are visible.
• Opposing advice is weighted 2-3 times more strongly than supportive advice.

Topics

• LLM Confidence Dynamics
• Choice-Supportive Bias
• Contradictory Information Overweighting
• Bayesian Decision-Making
• Model Calibration

Code references

• dharshsky/llm-confidence-biases

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• LLMs displaying less cognitive bias are not necessarily better decision makers — Large language models (LLMs) are known to include both social stereotypes and cognitive biases, which researchers are actively working to identify, characterize, and rectify.
• Quantifying Faithful Confidence Expression in Large Reasoning Models — Faithful confidence expression is a distinct, significant challenge for Large Reasoning Models, requiring new quantification methods.
• Do Emotions Influence Moral Judgment in Large Language Models? — LLMs' moral judgments are systematically biased by emotional context, unlike human judgments.
• CLEAR: Revealing How Noise and Ambiguity Degrade Reliability in LLMs for Medicine — LLMs exhibit a "humility deficit" in medical contexts, struggling with ambiguity and preferring incorrect answers over admitting uncertainty.
• The Same Email, Signed Differently: Testing Negotiation Bias and Recommendation Stability in LLMs — LLMs may exhibit negotiation bias and unstable recommendations based on sender identity cues.
• Motivated reasoning, confirmation bias, and AI risk theory — Confirmation bias, driven by motivated reasoning, significantly distorts beliefs in complex fields like AI risk by compounding across cognitive stages.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Nature Machine Intelligence.