AI Skill Report Card
Governing Ethical Decisions
Quick Start14 / 15
Given a proposal (feature, model, policy, or decision), produce three sections:
Recommendation▾
Add a brief note on how to handle ambiguous/insufficient information in the proposal (e.g., what to do if data lineage or stakeholder scope is unclear)
Ethical Assessment
[Identify ethical principles at stake and where they're upheld or violated]
Impact Analysis
[Who is affected, how, severity, likelihood, reversibility]
Recommendation
[Approve / Approve with conditions / Reject — with concrete required changes]
Example trigger: "Review this recommendation algorithm for ethical risks before launch."
Workflow14 / 15
Progress:
- Step 1: Understand the proposal — what it does, who builds it, who uses it, who is affected
- Step 2: Map against core ethical principles (see checklist below)
- Step 3: Identify stakeholders and analyze impact per group
- Step 4: Assess severity, likelihood, and reversibility of harms
- Step 5: Check for mitigations already in place vs. gaps
- Step 6: Formulate a clear, actionable recommendation
Step 2 — Ethical Principles Checklist
- Fairness/Bias: Does it treat groups equitably? Any proxy discrimination (age, gender, race, disability, socioeconomic status)?
- Privacy: What data is collected/used? Is consent informed? Is data minimized?
- Transparency: Can affected users understand how decisions are made? Is there disclosure?
- Accountability: Is there a clear owner/appeal path if something goes wrong?
- Autonomy: Does it manipulate, coerce, or reduce user agency (dark patterns, addictive design)?
- Safety/Harm: Physical, psychological, financial, reputational harm potential?
- Societal impact: Job displacement, misinformation, power concentration, environmental cost?
Step 3 — Stakeholder Impact Table
For each stakeholder group (end users, non-users affected, employees, vulnerable populations, society at large), note:
- Nature of impact (benefit/harm)
- Severity (low/medium/high/critical)
- Likelihood (rare/possible/likely/certain)
- Reversibility (easily reversible/costly to reverse/irreversible)
Recommendation▾
Include an example with a borderline/ambiguous case (not clearly approve or reject) to show nuanced judgment calls
Examples18 / 20
Example 1: Input: "We're building a hiring algorithm that scores resumes using past hiring data."
Output:
Recommendation▾
Consider adding a short section on how to weigh conflicting principles (e.g., transparency vs. privacy trade-offs) when they pull in opposite directions
Ethical Assessment
- Fairness/Bias: HIGH RISK. Training on historical hiring data risks encoding past discriminatory patterns (gender/school prestige/name-based bias).
- Transparency: Candidates have no visibility into scoring criteria — violates informed decision-making.
- Accountability: No stated appeal mechanism for rejected candidates.
Impact Analysis
| Stakeholder | Impact | Severity | Likelihood | Reversibility |
|---|---|---|---|---|
| Rejected candidates from underrepresented groups | Unfair exclusion from employment | Critical | Likely | Costly (lost opportunity) |
| Company | Legal/reputational liability (discrimination lawsuits) | High | Possible | Costly |
| Society | Reinforces structural inequality at scale | High | Likely | Hard to reverse once deployed widely |
Recommendation
REJECT in current form. Required before resubmission:
- Audit training data and model outputs for disparate impact across protected classes.
- Add human review for all rejections, not just approvals.
- Provide candidates a summary of key factors and an appeal channel.
- Re-test with fairness metrics (e.g., demographic parity, equalized odds) and set explicit thresholds.
**Example 2:**
Input: "A note-taking app wants to add a streak feature to boost daily engagement."
Output:
Ethical Assessment
- Autonomy: Streaks are a known engagement-manipulation pattern; low-to-moderate concern if not paired with loss-aversion tactics (e.g., punitive notifications).
- Safety/Harm: Minimal physical/financial harm; possible mild psychological pressure for compulsive users.
- Transparency: Low risk — mechanic is visible and understandable to users.
Impact Analysis
| Stakeholder | Impact | Severity | Likelihood | Reversibility |
|---|---|---|---|---|
| General users | Mild increase in habitual use | Low | Likely | Easily reversible (feature can be disabled) |
| Users prone to compulsive behavior | Anxiety over broken streaks | Medium | Possible | Easily reversible |
Recommendation
APPROVE WITH CONDITIONS:
- Avoid guilt-inducing copy ("Don't lose your streak!") — use neutral/encouraging language.
- Add a "streak freeze" or grace period to reduce anxiety-driven compulsive use.
- Make the feature opt-out.
Best Practices
- Always name the specific ethical principle(s) at stake — don't just say "this seems bad."
- Distinguish severity from likelihood; a rare catastrophic harm still requires mitigation.
- Prefer "Approve with conditions" over outright rejection when concrete fixes exist — be constructive.
- Consider second-order/societal effects, not just direct user impact.
- Flag irreversible harms as the highest priority regardless of likelihood.
- Keep language precise and avoid moralizing tone — this is a risk assessment, not a lecture.
Common Pitfalls
- Don't give vague verdicts like "seems fine" without mapping to specific principles.
- Don't ignore indirect/non-user stakeholders (society, competitors, environment).
- Don't treat "legal compliance" as equivalent to "ethical" — laws lag behind ethics.
- Don't recommend rejection without specifying what change would make it acceptable.
- Don't conflate low likelihood with low priority when severity is critical/irreversible.