🧠 human_ai_belief_dynamics

Modeling Belief Formation and Trust in Human–AI Interaction

[]

---

Overview

human_ai_belief_dynamics is a theory-driven dataset designed to study how humans form beliefs, assign trust, and make decisions in response to AI advice under uncertainty — particularly in high-stakes and vulnerable contexts.

Unlike conventional question–answer datasets, this dataset functions as an experimental stimulus and behavioral response set for Human–AI Interaction (HCI) and Human-Centered AI (HCAI) research. Each example models a controlled interaction between a user and an AI system, capturing belief priors, AI explanations, uncertainty framing, and resulting human responses.

The dataset includes theory-constrained synthetic behavioral responses, enabling early-stage experimentation and benchmarking prior to human-subject deployment.

---

Quick Start

from datasets import load_dataset

dataset = load_dataset("Dyra1204/human_ai_belief_dynamics")

---

Motivation

AI systems increasingly influence decisions in domains such as healthcare, finance, law, and employment. However, most benchmarks focus on model accuracy rather than how people interpret, trust, and act upon AI advice.

This dataset addresses three underexplored gaps in the literature:

1. Belief formation and updating — lack of benchmarks modeling how users revise beliefs in response to AI outputs
2. Trust miscalibration under uncertainty — limited study of trust collapse when AI systems express uncertainty or make errors
3. High-stakes and vulnerable decision contexts — underrepresentation of scenarios where the consequences of AI reliance are most significant

---

Dataset Design

Each record represents a single-turn human–AI interaction composed of:

• A decision scenario
• A user's prior belief
• An AI recommendation with explanation and uncertainty framing
• Human responses capturing belief update, trust, attribution, and decision

The dataset is constructed through a controlled experimental design with systematic manipulation of explanation style, uncertainty communication, and model correctness.

---

Schema

Each example contains the following variables:

Scenario Context

Field	Description
scenario_id	Unique identifier for the scenario
domain	Domain of the decision (e.g., healthcare, finance, legal)
stakes_level	low or high
scenario_text	Full text of the decision scenario presented to the user

AI Output

Field	Description
model_prediction	The AI system's recommendation or prediction
model_confidence	Expressed confidence level of the AI
explanation_style	One of: causal, statistical, narrative, none
explanation_text	The explanation provided alongside the prediction
uncertainty_framing	One of: explicit_prob, verbal_hedge, assertive, none
model_correctness	One of: correct, partial, incorrect

Human Response (Synthetic)

Field	Description
human_updated_belief	Posterior belief after receiving AI recommendation
human_trust_rating	Trust assigned to the AI system
perceived_model_competence	User's assessment of AI capability
perceived_model_transparency	User's assessment of AI clarity and openness
decision_taken	Final action or decision made by the user

---

Experimental Manipulations

The dataset systematically varies four independent factors across conditions:

Factor	Levels
Explanation Style	causal, statistical, narrative, none
Uncertainty Framing	explicit_prob, verbal_hedge, assertive, none
Model Correctness	correct, partial, incorrect
Stakes Level	low, high

Each base scenario is paired with multiple experimental conditions, enabling within-scenario comparisons across manipulations.

---

Intended Uses

This dataset is intended for:

• Human–AI interaction research
• Trust and reliance modeling
• Belief updating and anchoring studies
• Evaluation of explanation and uncertainty communication strategies
• Pre-human-subject experimental prototyping

---

Limitations

• Human responses are synthetic and theory-constrained, not collected from human subjects
• Interactions are single-turn only (no conversational memory or follow-up)
• Cultural, demographic, and individual differences are not modeled
• The dataset is designed to be replaceable or extensible with real human data once ethical approvals and study infrastructure are in place

---

Ethics Statement

• The dataset contains no real personal data
• All scenarios are fictional but plausible
• Vulnerable decision contexts are included strictly for research purposes, with no intent to provide real-world advice
• Users must not deploy this dataset for real decision support

---

Theoretical Grounding

The dataset design is grounded in established research across several disciplines:

• Trust in automation
• Algorithm aversion and appreciation
• Confidence and metacognitive judgment
• Explanation effects on trust and compliance
• Uncertainty communication in human–AI systems

Key influences include work by Hoffmann et al., Dietvorst et al., Yeomans et al., Koriat et al., and Ribeiro et al.

---

Citation

If you use this dataset in your research, please cite:

@dataset{human_ai_belief_dynamics_2026,
  author    = {Dyuti Dasmahapatra},
  title     = {human\_ai\_belief\_dynamics: Modeling Belief Formation and Trust in Human--AI Interaction},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Dyra1204/human_ai_belief_dynamics},
  doi       = {10.57967/hf/7827}
}

---

Author

Dyuti Dasmahapatra Hugging Face Profile