Responsible AI — Fairness and Bias

Google published its AI Principles in 2018, establishing seven objectives that guide all AI development at Google and on Google Cloud. These principles are directly referenced in MLE exam questions about responsible AI decisions.

Google also defined four AI application areas it will not pursue: technologies that cause overall harm, weapons, surveillance violating internationally accepted norms, and technologies that contravene international law and human rights.

Bias can enter an ML system at every stage: data collection, feature engineering, model training, evaluation, and deployment. Understanding the source of bias is the first step to mitigating it.

Data-Related Bias

Model and Evaluation Bias

There is no single definition of "fairness" — different definitions can even be mathematically incompatible. The right metric depends on the application context and the stakeholders affected.

Core Fairness Metrics

Detecting bias requires going beyond aggregate metrics. Slice-based evaluation is the most important technique: evaluate model performance separately for each demographic group.

Slice-Based Evaluation

Instead of reporting a single accuracy number, compute metrics per group: accuracy, precision, recall, F1, and AUC for each demographic slice. Look for significant disparities between groups. On Vertex AI, use the Model Evaluation feature to automatically compute slice-based metrics.

Confusion Matrix Per Group

Build a separate confusion matrix for each demographic group. Compare true positive rates (TPR), false positive rates (FPR), true negative rates (TNR), and false negative rates (FNR) across groups. Disparities in FPR are especially concerning in high-stakes applications (criminal justice, credit scoring).

Disparate Impact Analysis

Calculate the disparate impact ratio: the ratio of positive outcome rates between the unprivileged and privileged groups. A ratio below 0.8 (the "80% rule" from US employment law) indicates potential adverse impact and warrants investigation.

# Disparate impact calculation
def disparate_impact(y_pred, sensitive_attr, privileged_value, unprivileged_value):
    privileged_mask = sensitive_attr == privileged_value
    unprivileged_mask = sensitive_attr == unprivileged_value

    rate_privileged = y_pred[privileged_mask].mean()
    rate_unprivileged = y_pred[unprivileged_mask].mean()

    return rate_unprivileged / rate_privileged

# DI >= 0.8 passes the 80% rule
# DI < 0.8 indicates potential adverse impact

Mitigation techniques are categorized by when they are applied in the ML pipeline: before training (pre-processing), during training (in-processing), or after training (post-processing).

Pre-Processing Techniques

Modify the training data to reduce bias before the model ever sees it.

In-Processing Techniques

Modify the training algorithm itself to enforce fairness constraints.

Post-Processing Techniques

Adjust model outputs after training to satisfy fairness criteria.

The What-If Tool (WIT) is an interactive visual tool for exploring ML model behavior without writing code. It is available in TensorBoard, Jupyter/Colab, and Cloud AI Platform Notebooks.

Key Capabilities

# Launch What-If Tool in a Jupyter notebook
import witwidget
from witwidget.notebook.visualization import WitConfigBuilder, WitWidget

config = WitConfigBuilder(
    test_examples,          # Your test dataset
    feature_columns
).set_model_name("my-model")
 .set_target_feature("label")
 .set_label_vocab(["rejected", "approved"])

WitWidget(config)

Vertex AI integrates fairness into the ML lifecycle through Model Evaluation and Explainability features that operate directly in the managed platform.

Model Evaluation with Fairness Metrics

When you run a model evaluation job on Vertex AI, you can specify slicing specs to compute metrics per demographic group. The platform automatically generates confusion matrices, AUC-ROC curves, and precision-recall curves for each slice.

# Configure fairness slicing in Vertex AI evaluation
from google.cloud.aiplatform import ModelEvaluation

evaluation = model.evaluate(
    test_dataset=test_data,
    slicing_specs=[
        {"feature": "gender"},
        {"feature": "age_group"},
        {"feature": "ethnicity"}
    ]
)

Explainability

Vertex AI Explainable AI provides feature attributions that show which features contributed most to each prediction. This helps identify when sensitive attributes (or their proxies) are driving predictions. Methods include:

• Integrated Gradients — attribution method for neural networks
• XRAI — region-based attribution for images
• Sampled Shapley — model-agnostic attribution method

Documentation is a critical component of responsible AI. Data Cards and Model Cards are standardized templates for documenting the characteristics, limitations, and intended uses of datasets and models.

Model Cards

Introduced by Mitchell et al. (2019) at Google, a Model Card documents:

• Model Details — architecture, version, owner, date
• Intended Use — primary use cases, out-of-scope uses
• Factors — relevant demographic groups and environmental conditions
• Metrics — performance metrics chosen and why, per-group breakdowns
• Evaluation Data — datasets used for evaluation, their characteristics
• Ethical Considerations — known risks, limitations, sensitive use cases
• Caveats & Recommendations — known failure modes and mitigation advice

Data Cards

Data Cards document dataset characteristics: collection methodology, demographic composition, labeling process, known biases, and recommended uses. They help downstream users understand whether a dataset is appropriate for their use case.

AI regulation is evolving rapidly. ML engineers must be aware of the regulatory landscape to build compliant systems.

Responsible AI questions on the MLE exam test your ability to identify bias sources, choose appropriate fairness metrics, and select mitigation strategies for specific scenarios.

Identifying Bias Sources

• Model performs well overall but poorly for a specific group → Representation bias or evaluation bias
• Training data reflects historical inequities → Historical bias
• Different measurement standards across groups → Measurement bias
• Model used for a purpose it wasn't designed for → Deployment bias

Choosing Mitigation Strategy

Fairness Metric Selection

How to Explain This in 2 Minutes

Likely Interview Questions

Model Answers

System Design Scenario

Common Mistakes

• Relying on overall accuracy as proof of fairness — A model with 95% overall accuracy can have 99% accuracy for one group and 80% for another. Always compute sliced metrics per demographic subgroup. Overall metrics hide disparities.
• Removing sensitive attributes and assuming fairness — Simply dropping race or gender columns does not eliminate bias. Other features (zip code, school name, purchasing patterns) can serve as proxies. Fairness requires measuring outcomes across groups regardless of whether sensitive attributes are used as features.
• Treating fairness as a one-time check — Fairness can degrade over time due to data drift, population changes, or feedback loops. Production models need continuous fairness monitoring with automated alerts, not just a pre-launch review.