Model Cards

This section provides the basic identification and provenance of the model.

• Model Name & Version: Unique identifier for tracking and version control.
• Date: Creation or last major update date.
• Model Type: Architecture (e.g., BERT, ResNet-50, Gradient Boosting).
• Paper or Citation: Reference to the original research paper, if applicable.
• License: Terms of use (e.g., Apache 2.0, MIT, proprietary).
• Contact: Point of contact for questions or issues. This metadata is critical for reproducibility and responsible disclosure.

This section explicitly defines the primary purposes and appropriate contexts for model deployment.

• Primary Intended Uses: Specific tasks the model is designed for (e.g., sentiment analysis of product reviews, detecting pneumonia in chest X-rays).
• Primary Intended Users: The target audience (e.g., healthcare professionals, financial analysts, software developers).
• Out-of-Scope Uses: Clear warnings about misuse (e.g., "Not for diagnostic use without clinician review," "Not for evaluating loan eligibility"). This establishes the operational boundary and helps prevent misapplication of the model.

This is the quantitative core, reporting model efficacy across relevant metrics and datasets.

• Metrics: Task-appropriate measures (e.g., Accuracy, F1-score, BLEU, AUC-ROC).
• Evaluation Data: Description of the datasets used for testing, including their source and characteristics.
• Results: Aggregate performance scores. Crucially, this must include subgroup or slice-based analysis to surface disparities.
• Comparison: Benchmarks against relevant baselines or state-of-the-art models. This provides the empirical evidence for the model's capabilities and limitations.

This section documents a rigorous audit for potential biases and harms.

• Considered Sensitive Attributes: Lists protected attributes analyzed (e.g., race, gender, age), noting how they were constructed.
• Fairness Metrics: Reports results for chosen metrics (e.g., Demographic Parity, Equal Opportunity, Predictive Parity) across subgroups.
• Disparity Analysis: Highlights any significant performance gaps identified through subgroup or intersectional analysis.
• Known Limitations: Explicitly states any discovered biases, stereotypes, or fairness trade-offs. This fulfills the core transparency and accountability function of the Model Card.

This section details the data used for training and evaluation, as data provenance is a primary source of model behavior.

• Training Data: Description of the dataset(s), including size, source, collection methods, and any known biases (e.g., historical bias, representation bias).
• Evaluation Data: As above, for test/validation sets. Should be distinct from training data.
• Preprocessing: Notes on how data was cleaned, filtered, or transformed (e.g., tokenization, normalization, handling of missing values).
• Labeling Process: Description of how ground truth labels were generated, including annotator qualifications and agreement statistics. This transparency allows users to assess data suitability for their own context.

This section covers operational constraints, failure modes, and environmental factors.

• Hardware/Software: Inference requirements (e.g., GPU memory, specific libraries).
• Latency/Throughput: Expected inference speed under defined conditions.
• Known Failure Modes: Scenarios where the model performs poorly (e.g., on low-resolution images, domain-specific jargon, adversarial examples).
• Sensitivity: Notes on how predictions may change with small input perturbations.
• Environmental Impact: Estimated carbon footprint from training, if available. This information is essential for production deployment planning and risk assessment.

Model cards in regulated domains like healthcare adopt stringent templates to meet auditability requirements. They emphasize clinical validation and failure mode analysis.

• Key Sections: Intended patient population, contraindications, device specifications (for edge deployment), and clinical trial results.
• Performance Reporting: Metrics are broken down by clinically relevant subgroups (e.g., age, sex, ethnicity, disease subtype).
• References Standards: Often aligns with FDA's Software as a Medical Device (SaMD) pre-submission guidelines or ISO/IEC 24029-1 for AI assessment.
• Example: A model card for a diabetic retinopathy detection system would detail performance across different skin tones and camera types.

In finance, model cards are formalized into Model Risk Management (MRM) documentation, required by regulators (e.g., OCC, FRB). They focus on explainability, stability, and fairness.

• Documents challenge results: Includes outcomes from adversarial testing and drift detection system monitoring.
• Fairness Reporting: Mandates analysis under regulations like ECOA/Regulation B, reporting demographic parity and equal opportunity gaps for credit models.
• Third-Party Audit Trail: Includes sections for independent validation team sign-off, creating a clear audit trail for governance.
• Example: A card for an algorithmic trading model would detail its behavior during market stress scenarios.

The study and application of principles to ensure automated decision-making systems do not create unjust outcomes based on sensitive attributes. It provides the ethical framework that model cards operationalize by documenting performance across groups.

• Core Concern: Preventing discrimination in automated predictions.
• Relation to Model Cards: A model card's subgroup analysis section quantitatively reports on fairness metrics, translating abstract principles into measurable results.

A systematic, documented evaluation of an AI system to detect and measure potential discriminatory biases. A model card is the standardized reporting output of a comprehensive bias audit.

• Process: Involves subgroup analysis, fairness metric calculation, and review of training data.
• Output: The audit findings are summarized in the model card's 'Ethical Considerations' and 'Evaluation Results' sections, providing a snapshot of the system's fairness posture.

The practice of evaluating a model's performance metrics separately for distinct demographic or data slices. This is the primary analytical method required to populate a model card's fairness reporting.

• Purpose: To identify performance disparities (e.g., accuracy, FPR) masked by aggregate metrics.
• Model Card Implementation: Results are presented in tables or graphs comparing metrics like equal opportunity or demographic parity across groups defined by protected attributes.

A broader risk assessment process for deploying automated systems, often guided by policy. A model card serves as a core technical component within a full AIA, providing the empirical evidence on model behavior.

• Scope: Covers societal impact, regulatory compliance, and stakeholder consultation beyond pure model metrics.
• Synergy: The model card's documented limitations and intended use context are critical inputs for the AIA's risk evaluation and mitigation planning.

The degradation of a deployed model's fairness performance over time due to changing data. Model cards establish the baseline fairness profile against which future drift detection systems can monitor for bias drift.

• Cause: Shifting population statistics or evolving societal norms reflected in new data.
• Monitoring Link: The evaluation results in the model card provide the initial "fairness SLO" for continuous production monitoring, triggering card updates if significant drift is detected.