Summary

This PR adds a new evaluation metric suite, human_ai_trust, to Hugging Face Evaluate.

The goal of this metric is to support human-centered evaluation of AI systems by operationalizing:

• trust calibration
• belief updating
• uncertainty sensitivity
• asymmetric harm from overconfident errors
• explanation–confidence alignment

Unlike traditional metrics focused solely on predictive performance, this metric surfaces how users interpret, trust, and act on model outputs under uncertainty.

What's Included

This PR introduces:

• A new metric module: human_ai_trust
• Core metrics:
  • Expected Trust Error (ETE) — mean trust-confidence mismatch
  • Trust Sensitivity Index (TSI) — correlation between confidence and trust
  • Belief Shift Magnitude (BSM) — mean posterior-prior difference
  • Overconfidence Penalty (OCP) — weighted cost of confident errors
  • Normalized Overconfidence Penalty — bounded [0,1] variant
  • Explanation-Confidence Alignment (ECA) — correlation between complexity and confidence
• Full unit test coverage
• Comprehensive documentation
• An example notebook demonstrating real usage
• A companion reference dataset:
  • https://huggingface.co/datasets/Dyra1204/human_ai_trust_demo

Motivation

Accuracy alone is insufficient for evaluating AI systems in high-stakes or vulnerable contexts.

From a human–AI interaction perspective:

• a confident but wrong prediction is more damaging than a hesitant error
• users may over-trust highly confident systems (automation bias)
• users may ignore uncertainty signals
• beliefs may shift even when the model is wrong
• explanation style may distort trust

This metric suite provides theory-grounded, computational signals for evaluating these human-centered failure modes.

Design Philosophy

This metric is:

• Human-centered: grounded in cognitive science and HCI
• Descriptive, not causal: intended for evaluation, not inference
• Modular and composable: fits Hugging Face's metric API
• NaN-safe and edge-case robust
• Domain-agnostic: applicable across classification, QA, NLI, and generation

Related Work

This metric complements existing evaluation approaches:

• Traditional metrics (accuracy, F1, AUROC): measure predictive performance
• Calibration metrics (ECE, Brier score): measure probability alignment
• Human-AI trust metrics (this PR): measure human interpretation and decision-making under uncertainty

Key differences:

• Focuses on human beliefs rather than just model probabilities
• Captures asymmetric harm from overconfident errors
• Evaluates explanation-confidence alignment

Usage Example

import evaluate

metric = evaluate.load("human_ai_trust")

out = metric.compute(
    predictions=[1, 0, 1],
    references=[1, 1, 0],
    confidences=[0.9, 0.7, 0.8],
    human_trust_scores=[0.85, 0.6, 0.75],
    belief_priors=[0.3, 0.4, 0.5],
    belief_posteriors=[0.6, 0.5, 0.7],
    explanation_complexity=[10, 20, 15],
)

print(out)

Testing

All metrics are fully unit-tested with 100% code coverage, including:

• Zero-variance inputs
• Missing optional inputs
• Constant confidence values
• All-correct and all-wrong predictions
• Edge cases (empty arrays, single samples, etc.)

Run locally:

pytest metrics/human_ai_trust --cov=metrics/human_ai_trust

Companion Dataset

A small, theory-grounded demo dataset is available at:

https://huggingface.co/datasets/Dyra1204/human_ai_trust_demo

It demonstrates:

• trust calibration
• belief updating
• uncertainty communication
• explanation–confidence alignment

Limitations

These metrics are intended as descriptive human-centered evaluation signals.

They do not:

• replace task performance metrics
• infer causal effects
• evaluate explanation faithfulness
• measure fairness or bias
• substitute for real human-subjects data

Checklist

• New metric module added
• Full unit test coverage
• Documentation (README.md with formulas and examples)
• Example notebook
• Companion dataset published
• NaN-safe and edge-case handling
• Follows Hugging Face Evaluate API conventions
• Maintainer review

Feedback Welcome

I would very much welcome feedback on:

• API design
• metric naming
• default behaviors
• edge-case handling
• documentation clarity

I'm happy to adjust the implementation to better align with Hugging Face Evaluate conventions.