Ethics and Responsibility

Watch First

Learning Objectives

By the end of this lesson, you will be able to:

• Treat ethics as an engineering practice, not a public-relations layer.
• Identify risks around fairness, privacy, transparency, accountability, and harm.
• Run a simple group-level fairness check.
• Connect responsible AI to alignment, monitoring, deployment, and governance.

Responsible AI Loop

Responsible AI asks a simple but hard question:

Who is affected by this system, and are we building it in a way that deserves their trust?

In Flow Research-style systems, models may shape learning paths, community governance, public-good funding, or contributor reputation. That makes ethics part of the product requirements.

Launch Rule

Do not launch high-impact ML without a risk owner, monitoring plan, user recourse path, and rollback or fallback behavior.

Core Principles

Principle	Engineering question
Fairness	Does the system perform differently across groups?
Reliability and safety	Does it fail predictably and safely?
Privacy and security	Does it protect sensitive data?
Transparency	Can affected people understand the system's role?
Accountability	Who is responsible when harm occurs?
Human agency	Can people contest, override, or appeal decisions?

Principles matter only when they become concrete checks, docs, and operating procedures.

Stakeholder Mapping

Start by listing affected people and institutions.

For each group, ask:

• What benefit could they receive?
• What harm could they experience?
• What information do they need?
• How can they challenge or correct the system?

Fairness Metrics

Fairness is contextual. No single metric solves all cases, but group-level checks reveal hidden gaps.

Demographic parity difference compares positive prediction rates:

$$DPD = P(\hat{Y}=1|A=a) - P(\hat{Y}=1|A=b)$$

Equal opportunity difference compares true positive rates:

$$EOD = TPR_{A=a} - TPR_{A=b}$$

Use these as investigation tools, not automatic verdicts.

Fairness Check in Python

import pandas as pd
from sklearn.metrics import accuracy_score, recall_score

data = pd.DataFrame({
    "group": ["A", "A", "A", "B", "B", "B", "B", "A"],
    "y_true": [1, 0, 1, 1, 0, 1, 0, 0],
    "y_pred": [1, 0, 0, 1, 1, 1, 0, 0],
})

summary = []

for group, rows in data.groupby("group"):
    summary.append({
        "group": group,
        "n": len(rows),
        "accuracy": accuracy_score(rows["y_true"], rows["y_pred"]),
        "positive_prediction_rate": rows["y_pred"].mean(),
        "recall": recall_score(rows["y_true"], rows["y_pred"]),
    })

report = pd.DataFrame(summary)
print(report)

dpd = (
    report.loc[report["group"] == "A", "positive_prediction_rate"].iloc[0]
    - report.loc[report["group"] == "B", "positive_prediction_rate"].iloc[0]
)

eod = (
    report.loc[report["group"] == "A", "recall"].iloc[0]
    - report.loc[report["group"] == "B", "recall"].iloc[0]
)

print({"demographic_parity_difference": dpd, "equal_opportunity_difference": eod})

This does not prove fairness. It starts the investigation.

Sources of Harm

ML harms can appear at every stage.

Common risks:

• data collected without clear consent,
• sensitive attributes stored unnecessarily,
• proxies for protected attributes,
• labels reflecting past discrimination,
• no appeal process,
• opaque scoring,
• automation bias,
• models used outside intended context.

Privacy and Data Minimization

Responsible systems collect the least sensitive data needed for the task.

Ask:

• Do we need this field?
• Can it be aggregated or anonymized?
• Who can access it?
• How long should it be retained?
• Could it expose vulnerable users if leaked?

For learner-support systems, avoid collecting sensitive personal history unless it is clearly necessary, consented, protected, and governed.

Transparency and Recourse

A model does not need to expose every weight to be transparent. People need to understand:

• when a model is used,
• what decision it supports,
• what data categories influence it,
• what limitations exist,
• how to appeal or correct errors.

For high-impact systems, provide a human path. A learner should not be trapped by an automated label.

Responsible Deployment Checklist

Before launch, document:

• intended use,
• out-of-scope use,
• stakeholders,
• known risks,
• evaluation metrics,
• subgroup metrics,
• data sources,
• privacy controls,
• monitoring plan,
• fallback behavior,
• escalation owner,
• review schedule.

This can become a lightweight model card.

Ethics in CI/CD and Monitoring

Responsible AI should show up in engineering workflows.

Examples:

• fail a build if required data documentation is missing,
• warn if subgroup recall drops below a threshold,
• require review before deploying a model that affects user access,
• monitor fairness metrics over time.

Practical Exercises

Exercise 1: Stakeholder Map

Pick a model and map direct users, indirect stakeholders, operators, and decision makers.

Exercise 2: Run the Fairness Check

Run the Python example, then add another group or metric. Explain what the result suggests and what it does not prove.

Exercise 3: Write a Model Card Draft

Create a one-page model card with intended use, limitations, data, metrics, risks, and contact owner.

Self-Assessment

Rate yourself from 1 to 5:

• I can explain why ethics is part of ML engineering.
• I can identify fairness, privacy, transparency, and accountability risks.
• I can compute simple subgroup metrics.
• I can connect responsible AI to CI/CD, deployment, monitoring, and alignment.

Further Reading

• NIST AI Risk Management Framework
• Microsoft Responsible AI principles
• Fairlearn documentation
• Model Cards for Model Reporting
• Datasheets for Datasets

Next Steps

Use this lesson as a checklist for every advanced ML system you build. Capability without responsibility is not launch-ready.