Dyra1204/human_ai_trust_demo

Human–AI Trust & Belief Dynamics (Demo Dataset)

This dataset is a small, synthetic but theory-grounded benchmark designed to support human-centered evaluation of AI systems under uncertainty.

It accompanies the human_ai_trust metric in Hugging Face Evaluate.

---

What This Dataset Contains

Each row represents a single human–AI interaction instance with the following fields:

• prediction: model prediction (binary)
• reference: ground truth label
• confidence: model confidence in its prediction
• human_trust: human trust rating in the model output
• belief_prior: user's belief before seeing the AI output
• belief_posterior: user's belief after seeing the AI output
• explanation_length: proxy for explanation complexity

---

What This Dataset Is For

This dataset is intended to:

• Demonstrate the human_ai_trust evaluation metric
• Support research on:
  • trust calibration
  • belief updating
  • uncertainty communication
  • explanation–confidence alignment
• Provide a lightweight benchmark for HCI and HCAI experiments

---

How to Use

Install the datasets library if you haven't already:

pip install datasets

Load the dataset:

from datasets import load_dataset

ds = load_dataset("Dyra1204/human_ai_trust_demo")

Access individual fields:

predictions      = ds["train"]["prediction"]
references       = ds["train"]["reference"]
confidences      = ds["train"]["confidence"]
human_trust      = ds["train"]["human_trust"]
belief_prior     = ds["train"]["belief_prior"]
belief_posterior = ds["train"]["belief_posterior"]

Use it with the companion human_ai_trust metric:

import evaluate

metric = evaluate.load("human_ai_trust")

results = metric.compute(
    predictions=ds["train"]["prediction"],
    references=ds["train"]["reference"],
    confidence=ds["train"]["confidence"],
    human_trust=ds["train"]["human_trust"],
    belief_prior=ds["train"]["belief_prior"],
    belief_posterior=ds["train"]["belief_posterior"],
)

print(results)

---

What This Dataset Is Not

• It is not a real human-subjects dataset
• It is not suitable for training models
• It does not capture cultural, demographic, or contextual variation
• It does not reflect real medical, legal, or safety-critical decisions

---

How the Data Was Generated

The dataset was synthetically generated to reflect psychologically plausible dynamics:

• Model confidence is higher for correct predictions
• Human trust tracks confidence with noise
• Beliefs shift partially toward model confidence
• Explanations are longer when confidence is lower

This makes it suitable for exercising trust- and belief-based evaluation metrics without requiring human data collection.

---

Limitations

• Synthetic data cannot substitute for real behavioral data
• Trust and belief dynamics are simplified
• Explanation complexity is approximated via length
• No domain context is modeled

Users are encouraged to replace this dataset with real human-interaction data for empirical studies.

---

License

MIT