🤖 AI

Emulating Aggregate Human Choice Behavior and Biases with GPT Conversational Agents

This study demonstrates that GPT-4 and GPT-5 conversational agents can accurately emulate individual-level human cognitive biases and their interaction with contextual factors like cognitive load, while revealing distinct model-specific variations in aligning with human behavior.

Original authors: Stephen Pilli, Vivek Nallur

Published 2026-02-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Stephen Pilli, Vivek Nallur

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: Can AI Pretend to Be a Flawed Human?

Imagine you are trying to build a digital twin of a human. You want this digital twin to make decisions just like a real person. But here's the catch: real humans aren't perfect calculators. We get tired, we get confused, and we make irrational choices based on our feelings or habits. These irrational habits are called Cognitive Biases.

The most famous of these is the Status Quo Bias. It's like being so comfortable in your current armchair that you refuse to move to a better one, even if the new chair is more comfortable, just because you don't want to go through the effort of standing up.

The Question: Can a Large Language Model (LLM), like the AI powering this chat, not just act smart, but actually act like a flawed, biased human? And more importantly, can it do this when the conversation gets complicated and tiring (like when you have to remember a long list of numbers while making a choice)?

The Experiment: The "Tired Shopper" Test

The researchers set up a massive experiment with 1,100 real humans. They acted like a giant, digital shopping mall.

The Setup: The humans chatted with a friendly bot.
The Twist: Before the bot asked them to make a big decision (like choosing a job or investing money), it forced them to do one of two things:
- The "Easy Walk" (Simple Dialogue): The bot asked simple "Yes/No" questions about music or movies. (Low mental effort).
- The "Obstacle Course" (Complex Dialogue): The bot asked tricky questions where the humans had to remember a chain of facts about different artists or properties, doing math in their heads as they went. (High mental effort).
The Trap: After the walk or the obstacle course, the bot presented a choice. One option was framed as "what you currently have" (Status Quo), and the other was a new option.

The Human Result:

Yes, humans are biased: When the bot said, "You currently have Option A," people overwhelmingly stuck with Option A, even if Option B was slightly better.
The Tired Factor: Surprisingly, being tired from the "Obstacle Course" didn't make them more biased. They were just as stubbornly stuck in their ways whether they were fresh or exhausted.

The AI Test: Can the Bot Play Human?

Next, the researchers took the chat logs and the personal details (age, gender, location) of those 1,100 humans and fed them into different AI models (GPT-4 and GPT-5). They asked the AI: "Based on this person's chat history, what would they choose?"

They tested the AI with three different "personas":

The Neutral Robot: "Just answer the question."
The Natural Human: "Answer like a normal person would."
The Biased Human: "Answer like a human who makes irrational mistakes and is easily swayed by biases."

The Findings:

The AI is a Master of Mimicry: When the AI was told to act like a biased human (Persona #3), it predicted the human choices with about 68% accuracy. That's a very high score for a computer guessing human behavior!
The "Over-Acting" Problem: When the AI was told to be super biased, it sometimes went too far. It started seeing bias where humans didn't have any. It's like an actor who, when told to "cry," starts sobbing uncontrollably even when the scene only calls for a tear.
The Best Performer: The GPT-4.1-mini model was the star of the show. It was the best at predicting exactly how humans would behave without over-acting.

The Surprising Twist: The AI Didn't Need to "Read" the Chat

Here is the most fascinating part. The researchers tried to trick the AI. They took the chat logs and replaced the human's actual answers with random gibberish (like "banana, apple, 123").

Result: The AI's prediction accuracy barely dropped.

What this means: The AI wasn't really "reading" the human's personality or specific thoughts in the chat. Instead, it was relying on its massive training data to say, "Statistically, when humans face this specific type of choice, they usually pick the Status Quo."

It's like a magician who knows the audience will always pick the red card, so he doesn't need to read the audience's mind; he just knows the odds.

The Takeaway: What Does This Mean for Us?

AI is getting scary good at faking human flaws. We can use these AI models to simulate how thousands of people might react to a new policy, a new product, or a new law without needing to interview thousands of real people.
But be careful of the "Over-Acting." If you tell the AI to be biased, it might invent biases that don't exist. You have to be very careful with how you ask it to behave.
The "Tired" Factor didn't change the AI's mind. Just like the humans, the AI didn't get more biased when the task was hard. It just stuck to its statistical patterns.

The Bottom Line Analogy

Think of the AI as a method actor.

If you tell the actor, "Pretend to be a human making a choice," they will do a decent job.
If you tell them, "Pretend to be a human who is stubborn and lazy," they will do an excellent job, perhaps too excellent.
However, the actor isn't actually thinking about the specific person they are mimicking; they are just reciting lines from a script of "How Humans Behave" that they memorized from reading millions of books.

This paper proves that while AI can simulate human bias very well, it's doing so by recognizing patterns, not by truly "feeling" the confusion or the stubbornness of a tired human mind.

1. Problem Statement

While Large Language Models (LLMs) have demonstrated the ability to reproduce known cognitive biases in isolated, single-shot tasks, a critical gap remains in understanding their capacity to emulate dynamic, context-dependent human decision-making. Specifically, it is unclear whether LLMs can:

Predict individual-level biased decisions based on conversational history and demographic cues.
Reproduce the interaction between cognitive biases (specifically Status Quo Bias) and contextual factors (specifically Cognitive Load induced by dialogue complexity).
Faithfully simulate human irrationality in a conversational agent setting without explicit instruction to be biased, or if they require "over-alignment" prompts that distort natural behavior.

Current research often treats LLMs as static cognitive models, failing to account for how dialogue structure (e.g., memory demands, arithmetic complexity) influences the manifestation of bias in real-time interactions.

2. Methodology

The study employed a two-phase empirical approach involving a human baseline experiment followed by LLM simulation experiments.

Phase 1: Human Baseline Experiment

Participants: $N=1100$ recruited via Prolific.
Design: A $2 \times 3$ $2 \times 3$ between-subjects factorial design.
- Independent Variable 1 (Dialogue Complexity/Cognitive Load):
  - Simple Dialogue: Low-effort, binary preference questions (e.g., "Do you like pop music?").
  - Complex Dialogue: High cognitive load involving nested referential structures, arithmetic comparisons, and memory retention tasks (e.g., tracking attributes of multiple artists across turns).
- Independent Variable 2 (Decision Scenario Framing):
  - Three scenarios adapted from Samuelson & Zeckhauser (1988): Budget Allocation, Investment Portfolios, and College Job Selection.
  - Three framing conditions: Neutral, Status Quo A (Option A is current), Status Quo B (Option B is current).
Procedure: Participants interacted with a chatbot (GPT-4o Mini) via a Streamlit interface. They completed a prior dialogue task (Simple or Complex) followed by a decision scenario.
Metrics: Decision outcomes, NASA-TLX (self-reported cognitive load), recall accuracy, and response times.

Phase 2: LLM Agent Experiments

Models Tested: GPT-4.1, GPT-4.1-mini, GPT-5, and GPT-5-mini.
Simulation Setup: LLM agents were provided with:
1. Demographic data of the human participant.
2. The full transcript of the human's prior dialogue (including the complex/simple interaction).
3. The decision scenario.
Human-Likeness Prompting Levels (HL):
- HL1 (Minimal): "You are a participant in a research study."
- HL2 (Naturalistic): "Answer questions as naturally as you would in everyday life."
- HL3 (Explicit Bias): "Act as a human prone to cognitive biases... lean into intuitive, sometimes irrational judgments."
Evaluation Metrics:
- Individual-Level Precision: Accuracy of the LLM in predicting the specific choice made by the human participant.
- Sample-Level Precision: Ability to reproduce the presence or absence of bias (True Positive/Negative rates).
- Mean Absolute Difference (MD): Alignment of the change in effect size between Simple and Complex dialogue conditions (capturing the interaction effect).

3. Key Contributions

Context-Aware Bias Simulation: The paper moves beyond static bias detection to investigate how LLMs handle the interaction between cognitive load and bias. It establishes that while LLMs can reproduce biases, their ability to simulate the modulation of bias by context varies significantly by model and prompting strategy.
Human Baseline in Dialogue: It provides a rigorous, reproducible human baseline for Status Quo Bias within a conversational agent context, confirming that chatbots effectively elicit this bias and that complex dialogue increases perceived cognitive load (though it did not statistically increase the bias magnitude in humans).
Prompt Sensitivity Analysis: The study reveals a trade-off in prompting:
- Neutral prompts (HL1/HL2) yield high precision in reproducing the presence of bias (matching human data) but fail to capture the interaction with cognitive load.
- Explicit bias prompts (HL3) force the model to simulate bias even in scenarios where humans showed none (false positives), leading to lower precision, though they sometimes better captured the direction of interaction effects.
Individual vs. Aggregate Prediction: The study demonstrates that while LLMs can accurately predict aggregate human tendencies (sample-level), their ability to predict individual decisions relies more on general task patterns than on specific individual cues from the dialogue transcript.

4. Key Results

Human Results

Status Quo Bias Confirmed: Participants exhibited a strong Status Quo bias in Budget Allocation and College Job scenarios (e.g., selection rates for the status quo option increased significantly when framed as such).
Cognitive Load Validation: Complex dialogue significantly increased NASA-TLX scores (Mental Demand and Effort) and response times, confirming the manipulation was successful.
Interaction Effect: Contrary to some hypotheses, increased cognitive load did not significantly strengthen the Status Quo bias in human participants across the tested scenarios. The bias remained robust but stable regardless of dialogue complexity.

LLM Results

Individual Prediction: GPT-4.1 achieved the highest individual-level prediction precision (~68.5% under HL2), significantly above random chance. However, perturbation tests (replacing human dialogue with random text) showed only a minor drop in precision, suggesting LLMs rely on general task patterns rather than specific individual cues.
Sample-Level Bias Reproduction:
- HL1 & HL2: Achieved perfect precision (1.0) in reproducing the existence of bias (matching human results where bias exists and where it doesn't).
- HL3: Showed lower precision (0.67) due to "over-alignment," predicting bias in scenarios (like Investment Portfolios) where humans showed none.
Contextual Interaction (The Critical Finding):
- Humans showed no significant change in bias magnitude between Simple and Complex dialogue.
- HL1/HL2 Models: Failed to align with this human pattern; they showed divergent effect sizes between conditions (poor MD scores).
- HL3 Models: Showed a pattern closer to humans regarding the direction of change (e.g., increasing bias under load in College Jobs), but this was often an artifact of forcing bias everywhere.
- GPT-4.1-mini (HL2): Showed the most promising alignment, capturing both the presence of bias and a closer approximation of the interaction dynamics compared to other models.

5. Significance and Implications

For HCI and AI Design: The findings suggest that LLMs can serve as effective proxies for simulating human decision-making in policy and experimental contexts, but prompt engineering is critical. Explicitly instructing models to be "biased" leads to unrealistic over-sensitivity, while neutral prompts yield high fidelity for aggregate trends but may miss subtle context-dependent shifts.
Limitations of Current LLMs: LLMs currently struggle to simulate the nuanced interaction between cognitive load and bias without explicit, often distorting, instructions. They appear to model "what humans tend to do" (aggregate statistics) rather than "what this specific human would do" (individual reasoning) in complex conversational flows.
Future Directions: The study highlights the need for adaptive, bias-aware AI systems that can detect high cognitive load in users and simplify interactions to mitigate bias, rather than relying on LLMs to perfectly simulate human irrationality in all contexts. It also calls for further research into other biases and more realistic, domain-specific dialogue scenarios beyond abstract text-based tasks.

In conclusion, the paper establishes that while LLMs are powerful tools for emulating aggregate human biases, their ability to faithfully replicate the dynamic interplay between context (cognitive load) and bias is still evolving and highly dependent on the specific model architecture and prompting strategy used.