From Expectation to Evaluation: Expectation Cues Systematically Bias LLM and Human Judgment

Paper Title

From Expectation to Evaluation: Expectation Cues Systematically Bias LLM and Human Judgment

Publication Info

• Topic area: The influence of expectation cues on evaluative judgments in humans and large language models (LLMs).
• Keywords: Expectation cues, judgment bias, LLM evaluation, human-AI interaction, decision support, cognitive bias, priming effects, fairness, transparency, evaluation distortion.

Background and Problem

• Problem / challenge: Prior research has shown that expectation cues (e.g., source labels, expertise signals) bias human judgment, but it remains unclear whether LLMs exhibit similar biases and how these biases compare to human patterns.
• Significance: Understanding expectation-driven biases in LLMs is critical for ensuring fairness, reliability, and transparency in high-stakes domains like healthcare, education, and law.
• Motivation and related work: While LLMs have been shown to respond to contextual signals like tone and politeness, their sensitivity to expectation cues and the structural similarities to human biases remain unexplored. This study addresses this gap by systematically comparing humans and LLMs under controlled conditions.

Solution

• Proposed approach: A systematic investigation of expectation-driven biases in humans and LLMs through two experiments and an additional Chain-of-Thought (CoT) reasoning condition.
• Novelty:
  1. Development of a matched experimental paradigm to compare human and LLM evaluative judgments under identical conditions.
  2. Introduction of expectation cues as contextual priors influencing evaluation, extending classical cognitive theories to artificial systems.
  3. Empirical evidence of shared and distinct judgment adjustment patterns between humans and LLMs.
  4. Exploration of CoT reasoning to assess its impact on mitigating expectation-driven biases in LLMs.
• Procedure and key techniques:
  1. Two experiments with factorial designs manipulating expectation levels (high, moderate, low, none) and suggestion quality (high, moderate, low).
  2. Experiment 1: Evaluators judged suggestions after receiving expectation cues.
  3. Experiment 2: Evaluators re-evaluated suggestions after receiving expectation cues between two evaluations.
  4. Additional experiment: CoT reasoning was introduced to test its effect on bias.
  5. Comparison of human evaluators and three LLMs (GPT-4o, Llama-3.3-70B, DeepSeek-r1) using standardized prompts and evaluation metrics.

Results

• Concrete findings:
  1. Expectation cues systematically biased evaluative judgments in both humans and LLMs, with higher expectations leading to more favorable evaluations.
  2. Larger mismatches between expectation and suggestion quality amplified judgment distortions.
  3. Humans showed asymmetric resistance to negative expectation violations, while LLMs adjusted symmetrically to both positive and negative violations.
  4. CoT reasoning did not mitigate expectation-driven biases in LLMs.
  5. Humans exhibited discrepancies between stated intentions and actual judgment changes, while LLMs showed complete alignment between reported and observed outputs.
• Advantage over baselines: The study provides the first systematic comparison of expectation-driven biases in humans and LLMs, revealing structural similarities and differences in judgment patterns.
• Experiments / evaluation:
  • Experiment 1: 720 evaluations per modality (humans and three LLMs) across 12 conditions.
  • Experiment 2: 540 evaluations per modality with a two-stage evaluation design.
  • Additional experiment: CoT reasoning tested with the same LLMs and conditions.
  • Metrics: Judgment Score (J), Expectation Scale (E), and judgment alteration (ΔJ).
• Limitations and future work:
  1. Limited depth of human evaluation due to crowdsourced settings.
  2. Short-term responses in controlled environments may not generalize to real-world, long-term interactions.
  3. Monolingual, text-only context limits applicability to multimodal or culturally diverse settings.
  4. Future work should explore mitigation strategies like expectation-calibration interfaces and neutral framing templates.

Summary

This study demonstrates that expectation cues systematically bias evaluative judgments in both humans and LLMs, with higher expectations leading to more favorable evaluations and larger mismatches amplifying distortions. Humans and LLMs exhibit both shared and distinct adjustment patterns, with LLMs showing consistent and symmetric responses to expectation violations. Chain-of-Thought reasoning did not mitigate biases in LLMs. These findings highlight the importance of considering expectation cues in designing fair and transparent human–AI interactions and suggest opportunities for developing bias mitigation strategies in AI systems.