Performance Metrics and Limitations
What you'll learn: How to honestly report your model's performance across multiple metrics and datasets while transparently documenting where it fails.
The Honest Scorecard
Reporting a single accuracy number on your training set is like a restaurant claiming "best food" without specifying who judged it or what cuisine. Real-world ML deployment demands comprehensive performance documentation across:
- Multiple metrics: accuracy, precision, recall, F1, AUC-ROC for classification; MAE, RMSE for regression
- Diverse datasets: training, validation, test, and real-world performance
- Demographic subgroups: performance broken down by race, gender, age, etc.
- Edge cases: how the model behaves on rare or unusual inputs
Why This Matters for Governance
Stakeholders need to understand not just where your model excels, but where it doesn't. A hiring model might show 92% accuracy overall but only 78% for candidates over 50. A medical diagnosis system might perform well on common conditions but catastrophically fail on rare diseases.
Document known limitations explicitly: Does your model struggle with non-English text? Does performance degrade after 6 months? Does it require high-quality images? These aren't embarrassments—they're critical information for safe deployment.
Failure Modes Are Features
Every model has failure modes—specific conditions under which it produces unreliable outputs. Maybe your sentiment analyzer fails on sarcasm, or your object detector misclassifies certain backgrounds. Documenting these in your model card prevents downstream misuse and builds trust through transparency.
Think of limitations documentation as a medication's side effects label: it doesn't mean the treatment is bad, but users deserve to know the risks before relying on it.
Key Takeaway: Comprehensive performance reporting across metrics, datasets, and demographic groups—paired with honest documentation of limitations and failure modes—is essential for responsible AI deployment and regulatory compliance.