Large language models are increasingly used in decision-making tasks that require them to process information from a variety of sources, including both human experts and other algorithmic agents. How do LLMs weigh the information provided by these different sources? We consider the well-studied phenomenon of algorithm aversion, in which human decision-makers exhibit bias against predictions from algorithms. Drawing upon experimental paradigms from behavioural economics, we evaluate how eightdifferent LLMs delegate decision-making tasks when the delegatee is framed as a human expert or an algorithmic agent. To be inclusive of different evaluation formats, we conduct our study with two task presentations: stated preferences, modeled through direct queries about trust towards either agent, and revealed preferences, modeled through providing in-context examples of the performance of both agents. When prompted to rate the trustworthiness of human experts and algorithms across diverse tasks, LLMs give higher ratings to the human expert, which correlates with prior results from human respondents. However, when shown the performance of a human expert and an algorithm and asked to place an incentivized bet between the two, LLMs disproportionately choose the algorithm, even when it performs demonstrably worse. These discrepant results suggest that LLMs may encode inconsistent biases towards humans and algorithms, which need to be carefully considered when they are deployed in high-stakes scenarios. Furthermore, we discuss the sensitivity of LLMs to task presentation formats that should be broadly scrutinized in evaluation robustness for AI safety.
Figures & tables
Figure 1: ‘Stated’ algorithm aversion across all models and tasks, operationalized as the gap between the trust rating given to the human expert and the algorithm.
Figure 2: Aggregate probabilities that LLMs in Study 2 demonstrate in delegating an algorithmic agent or a human expert to make the next prediction in the task, or neutral (no indicated preference).
Figure 3: Probability that each LLM correctly bets on the stronger predictor, disaggregated by task and whether the stronger predictor is presented as a human expert or an algorithm.
Variable
Coefficient
p -Value
Intercept
β=−0.78
p<0.001
Strong Algorithm
β=2.04
p<0.001
Complex LLM
β=1.52
p<0.001
Interaction
β=0.58
p<0.001
Table 1: Regression coefficients and p -values for a mixed effects logistic model predicting whether LLMs will delegate the task to the more accurate agent.
Figure 4: Probability of choosing the human expert over the algorithm in Studies 1 and 2, demonstrating the stated-revealed trust inconsistency. Error bars are SEM.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Task
Human Expert
Task
Human Expert
Estimating air traffic †
Air traffic controller
Recommending a marketing strategy
Marketer
Piloting a plane
Pilot
Predicting student performance †
Admissions officer
Diagnosing a disease
Cardiologist
Predicting employee performance
Manager
Giving directions
Navigator
Hiring and firing employees
Manager
Analyzing data
Analyst
Playing a piano
Musician
Driving a subway
Conductor
Writing news article
Journalist
Appendix
Table C.1: List of tasks used in Study 1 from Castelo et al. , ordered in descending task objectivity (as ranked in the original study). We designate a corresponding human expert role for each task. Bolded tasks are a representative subset of tasks included in Study 2, where † indicates that the task was also in Dietvorst et al. .
Figure D.1: Correlation between the LLMs trust gaps.
Figure D.2: Regression coefficients between LLMs’ responses ( y-axis ) and human responses from Castelo et al. ( x-axis ) with the gap in trust between human experts and algorithms. LLM-rated gaps with high statistical significance of p<0.001 are outlined in black with stronger color saturation. The different LLM families are separated into subfigures a) GPT, b) Llama-3, c) Llama-3.1, and d) Claude.
Figure D.3: Robustness experiments conducted to test the prompt variation of Study 1, where the algorithm is framed as (a) an LLM agent or (b) an expert algorithm .
Variable
Coefficient
p -Value
Intercept
β=0.74
p<0.001
Complex LLM
β=−1.77
p<0.001
Appendix
Table E.1: Regression coefficients and p -values for a mixed effects logistic model predicting whether LLMs will incorrectly bet on an algorithm when shown a stronger human alternative.
Figure E.1: Baseline version of Figure 3 with two algorithmic predictors instead of a human and an algorithm. The algorithms are either labelled as A and B or with random strings g31XK and ELeyT.
Figure F.1: Replication of the Study 1 results in Figure 1 with newer models from OpenAI, Anthropic, and Meta in 2026.
Figure F.2: Updated Figure 2 with newer models from OpenAI, Anthropic, and Meta in 2026.
Variable
Coefficient
p -Value
Intercept
β=2.65
p<0.001
Strong Algorithm
β=1.46
p=0.06
Complex LLM
β=0.66
p=0.02
Interaction
β=0.67
p=0.33
Appendix
Table F.1: Updated regression from Table 1 using responses from newer models in 2026.
Figure F.3: Replication of the stated-revealed comparison results in Figure 4 with newer models.
Large language models (LLMs) are increasingly used as automated evaluators (LLM-as-a-Judge). This work challenges its reliability by showing that trust judgments by LLMs are biased by disclosed source labels. Using a counterfactual design, we find that both humans and LLM judges assign higher trust to information labeled as human-authored than to the same content labeled as AI-generated. Eye-tracking data reveal that humans rely heavily on source labels as heuristic cues for judgments. We analyze LLM internal states during judgment. Across label conditions, models allocate denser attention to the label region than the content region, and this label dominance is stronger under Human labels than AI labels, consistent with the human gaze patterns. Besides, decision uncertainty measured by logits is higher under AI labels than Human labels. These results indicate that the source label is a salient heuristic cue for both humans and LLMs. It raises validity concerns for label-sensitive LLM-as-a-Judge evaluation, and we cautiously raise that aligning models with human preferences may propagate human heuristic reliance into models, motivating debiased evaluation and alignment.
Xin Sun, Di Wu, Sijing Qin +3
National Institute of Informatics (NII), Japan · University of Amsterdam, the Netherlands · Hitotsubashi University, Japan +3
Current Large Language Models (LLMs) are gradually exploited in practically valuable agentic workflows such as Deep Research, E-commerce recommendation, and job recruitment. In these applications, LLMs need to select some optimal solutions from massive candidates, which we term as LLM-as-a-Recommender paradigm. However, the reliability of using LLM agents for recommendations is underexplored. In this work, we introduce a BiasRecommendation Benchmark (BiasRecBench) to highlight the critical vulnerability of such agents to biases in high-value real-world tasks. The benchmark includes three practical domains: paper review, e-commerce, and job recruitment. We construct a Bias Synthesis Pipeline with Calibrated Quality Margins that 1) synthesizes evaluation data by controlling the quality gap between optimal and sub-optimal options to provide a calibrated testbed to elicit the vulnerability to biases; 2) injects contextual biases that are logical and suitable for option contexts. Extensive experiments on both SOTA (Gemini-{2.5,3}-pro, GPT-4o, DeepSeek-R1) and small-scale LLMs reveal that agents frequently succumb to injected biases despite having sufficient reasoning capabilities to identify the ground truth. These findings expose a significant reliability bottleneck in current agentic workflows, calling for specialized alignment strategies for LLM-as-a-Recommender. The complete code and evaluation datasets will be made publicly available shortly.
Zichen Tang, Zirui Zhang, Qian Wang +3
The Hong Kong University of Science and Technology · National University of Singapore · The Hong Kong University of Science and Technology (Guangzhou)
We test whether LLMs show robust decision biases. Treating models as participants in two-arm bandits, we ran 20000 trials per condition across four decoding configurations. Under symmetric rewards, models amplified positional order into stubborn one-arm policies. Under asymmetric rewards, they exploited rigidly yet underperformed an oracle and rarely re-checked. The observed patterns were consistent across manipulations of temperature and top-p, with top-k held at the provider default, indicating that the qualitative behaviours are robust to the two decoding knobs typically available to practitioners. Crucially, moving beyond descriptive metrics to computational modelling, a hierarchical Rescorla-Wagner-softmax fit revealed the underlying strategies: low learning rates and very high inverse temperatures, which together explain both noise-to-bias amplification and rigid exploitation. These results position minimal bandits as a tractable probe of LLM decision tendencies and motivate hypotheses about how such biases could shape human-AI interaction.
Haomiaomiao Wang, Tomás E Ward, Lili Zhang
Insight Research Ireland Centre for Data Analytics, Ireland · School of Computing, Dublin City University, Ireland