Research map / Large language models / AI and society
Large Language Models Assume People are More Rational than We Really are
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
The paper shows that LLMs (GPT-4o, GPT-4-Turbo, Llama-3, Claude 3 Opus) assume people are more rational than they are, in both forward (choice simulation/prediction on choices13k) and inverse (inferring preferences from choices) decision-making, with inverse inferences correlating strongly with human inferences.
Problem
Whether LLMs' implicit models of human decision-making match how people actually make choices, or instead assume people are more rational (expected-value maximizing) than they are.
Approach
Uses two psychology paradigms, risky choice (forward modeling on choices13k) and preference inference from observed choices (inverse modeling on the Jern et al. 47-decision task), to show LLMs align with expected-value rational models when predicting/simulating choices, yet their inverse inferences match human inferences, which themselves assume rationality.
Key results
- GPT-4o with CoT correlates 0.94 with maximum expected value on individual risky choice prediction, versus 0.48 for humans.
- Inverse modeling: GPT-4o CoT rankings correlate 0.95 (absolute utility) and 0.94 (relative utility) with rational models in the positive context.
- GPT-4o CoT inverse rankings correlate 0.97 Spearman with human inferences in the positive context and 0.87 in the negative context.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CoT instead of zero-shot prompting | GPT-4o predicting individual human choices on choices13k · choices13k, individual human choice | Spearman correlation with maximum expected value: 0.2843 -> 0.9444 | Helped |
| Use CoT instead of zero-shot prompting | GPT-4o predicting individual human choices on choices13k · choices13k, individual human choice | Spearman correlation with human choices: 0.4756 -> 0.6113 | Helped |
| Predict the aggregate choice distribution rather than individual choices (ablation) | GPT-4o, zero-shot forward modeling · choices13k, human choice prediction | Spearman correlation with maximum expected value: 0.2843 -> 0.1741 | Hurt |
| Simulate a participant rather than predict an individual human choice, using CoT (ablation) | GPT-4o forward modeling · choices13k, maximum expected value | Spearman correlation with maximum expected value: 0.9444 -> 0.9255 | No clear effect |
| Increase model capability from Llama-3-8B to GPT-4o with CoT | Inverse modeling, positive context · Inverse modeling, positive (candies) | Spearman correlation with absolute utility: 0.62 -> 0.95 | Helped |
| Use CoT instead of zero-shot prompting | GPT-4o inverse modeling · Inverse modeling, positive (candies) | Spearman correlation with absolute utility: 0.59 -> 0.95 | Helped |
| Use CoT instead of zero-shot prompting | GPT-4o inverse modeling · Inverse modeling, negative (shocks) | Spearman correlation with human inferences: 0.51 -> 0.87 | Helped |
| Vary GPT-4o temperature from 0.0 to 1.5 in CoT prompting (ablation) | GPT-4o forward modeling · choices13k, maximum expected value | Spearman correlation with maximum expected value: 0.540 -> 0.351 | Hurt |
| Predict a monkey's choices rather than a human's, with CoT (ablation) | GPT-4o forward modeling · choices13k, maximum expected value | Spearman correlation with maximum expected value: 0.970 -> 0.867 | Hurt |
| Predict a monkey's choices rather than a human's, with CoT (ablation) | GPT-4o forward modeling · choices13k, human choices | Spearman correlation with human choices: 0.549 -> 0.630 | Helped |
| Vary the demographic profile of the decision maker (ablation) | GPT-4o forward modeling with CoT · choices13k, maximum expected value | Spearman correlation with maximum expected value: 0.964; 0.965; 0.969; 0.971; 0.972 | No clear effect |
| Use a behavioral model instead of the other fitted models to explain GPT-4 zero-shot responses | GPT-4 individual zero-shot predictions · choices13k, individual human choice | MSE with GPT-4 outputs: 0.11427 -> 0.09835 | Helped |
Surprising findings
- Prompting GPT-4o to simulate monkeys yields less rational predictions that correlate more with actual human choices (CoT r=0.63 vs human).
- Zero-shot GPT-4-Turbo predictions sometimes ignore probabilities, best fit by a mixture of theories with a flat probability weighting function.
Limitations
- Abstract, controlled domains may not generalize to real-world contexts.
- No direct comparison with human predictions of others' choices on the same decisions.
- Simple demographic-based simulation; choices13k lacks subject-level details.
Experimental setup
Datasets: choices13k Jern et al. 47-decision inverse modeling human data
Benchmarks: choices13k Jern et al. inverse decision-making task
Metrics: Spearman correlation Pearson correlation MSE
Baselines: Expected value (rational model) Absolute utility model Relative utility model Likelihood model Marginal likelihood model Prospect Theory Expected Utility
Builds on: choices13k Jern et al. inverse decision-making model Mixture of Theories chain-of-thought prompting
Closest papers by meaning
- Language Models Trained to do Arithmetic Predict Human Risky and Intertemporal Choice (ICLR 2025 · similarity 0.90)
- Decision-Making Behavior Evaluation Framework for LLMs under Uncertain Context (NeurIPS 2024 · similarity 0.88)
- LLM Agents Can Be Choice-Supportive Biased Evaluators: An Empirical Study (AAAI 2025 · similarity 0.88)
- Are Large Language Models Sensitive to the Motives Behind Communication? (NeurIPS 2025 · similarity 0.88)
- Distributive Fairness in Large Language Models: Evaluating Alignment with Human Values (NeurIPS 2025 · similarity 0.87)
- Turning large language models into cognitive models (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.