Lesson 5.10 - Ethics & Limitations of ML in Engineering

Target: ~10 min read - 15 min discussion

Overview

A discussion-based closing lesson. Three themes: data bias (a model trained on unrepresentative data makes unreliable predictions for underrepresented conditions - e.g. a failure model trained on one equipment brand may not generalize to another); model explainability (a black-box prediction is harder to defend in a design review or regulatory submission than a transparent calculation - which matters more in engineering than in many other ML domains); and knowing when to trust a model - statistical models describe correlations in historical data; engineering judgment, physical reasoning, and applicable codes still govern final decisions.

Why This Matters (PH Context)

The PRC and the professional bodies (PICE, IIEE, PSME, PIChE, and others) hold the licensed engineer of record personally accountable for a design decision - "the model said so" is never a substitute for engineering judgment, and every model-informed decision in a regulatory or design context should trace back to a professional's sign-off.

Discussion Prompts (no code - reflect and write short answers)

1. Data bias: the equipment-failure model in Lesson 5.7 was trained on synthetic data with a specific relationship between vibration, temperature, hours, and failure. If a real dataset came predominantly from one manufacturer's machines, what could go wrong applying that model to a different manufacturer's equipment?

2. Explainability trade-off: the random forest slightly outperformed the single decision tree, but the tree is far easier to explain to a non-technical stakeholder. In a regulatory submission or design review, when would you choose the more interpretable but slightly less accurate model?

3. Regulatory context: how does the PRC currently treat AI/ML-assisted engineering analysis - is there a formal framework yet? What does that imply about how much weight a licensed engineer should place on a model's output vs. their own calculations?

4. When NOT to use ML: describe one problem from your own work where a well-established physical model (a code equation, a conservation law) is clearly more appropriate than a data-driven model - and why.

Suggested Format

A short (150-300 word) written reflection per prompt, submitted as a discussion post rather than a graded exercise.

---