Bias taught you to measure slice gaps after Vol. 19 quality scores. Fairness is the design choice: which notion of equality the product commits to, who signs it, and what you are willing to trade against utility. There is no single metric that is “the fair one” for every loan, chatbot, or hiring screen.
This lecture sits between measurement (bias) and accountability artifacts (explainability, transparency, governance, compliance). It is educational criteria literacy—not legal advice and not a fake empirical study of any protected class.
Learning Objectives
By the end of this lesson, students should be able to:
- State fairness as a goal/policy choice built on bias measurements, not a substitute for them.
- Contrast group criteria (parity, equalized odds/opportunity) with individual/similarity notions.
- Explain why common group criteria can conflict (impossibility at a high level).
- Compute selection rates and a simple equalized-odds sketch on toy labels only.
- Document trade-offs with accuracy/utility and name an owner for the chosen definition.
- Link fairness choices to transparency, privacy (attribute collection), and governance.
Fairness (in this module) is an explicit, stakeholder-owned criterion for how benefits, errors, or opportunities may differ across people or groups in a defined use case. It is not “the model feels nice,” and it is not automatically “equal accuracy.” You first measure bias (gaps). Then you choose a fairness target—e.g. similar selection rates, similar error rates among qualified cases, or similar treatment of similar individuals—knowing that targets can disagree. Jurisdiction, product risk, and compliance constrain the menu; this lecture does not declare what the law requires.
Group Criteria You Will See in Papers and Reviews
| Criterion (concept) | Informal claim | Watch-out |
|---|---|---|
| Demographic / statistical parity | Positive decision rates are similar across groups. | Ignores base rates and “qualified” differences; can conflict with accuracy. |
| Equal opportunity | Among true positives (e.g. actually qualified), recall is similar. | Needs a trustworthy label of “qualified.” |
| Equalized odds | TPR and FPR are both similar across groups. | Stricter than equal opportunity; often tighter utility trade-off. |
| Calibration by group | A score of 0.8 means ~80% in each slice. | Can conflict with equalized odds under typical assumptions. |
| Individual / similarity | Similar people (on allowed features) get similar outcomes. | Depends entirely on the similarity metric—which can itself be biased. |
These are mathematical templates. They do not tell you which protected or operational slices are lawful to use, whether the label is just, or whether a chatbot’s stereotype problem is even a binary decision problem. Generative products often need content and representation tests (Vol. 19 human evaluation + bias slices) rather than only TPR/FPR.
Why You Cannot “Satisfy All Fairness Metrics”
Impossibility (high level)
- If groups have different base rates and the predictor is imperfect, parity, equalized odds, and calibration generally cannot all hold.
- Picking one criterion is an ethical/product decision, not a math bug.
- Report which criterion you chose and which you knowingly dropped.
Label justice
- Equalized odds assumes \(y\) is the right target.
- Historical labels can encode the unfair process you hoped to fix.
- Sometimes the fair move is to change the task, not reweight the model.
Utility trade-off
- Constraints usually cost some global accuracy or profit.
- That cost can still be the right call for risk and trust.
- Hide the trade-off and reviewers will rediscover it in production.
Explicit fairness goals buy
- A testable acceptance bar beside Vol. 19 metrics
- Shared language for product, ML, and legal review
- A place to record trade-offs on a model card
Toy Policy + Metric Sketch
Again: synthetic groups only. The code shows how an engineer would implement a chosen criterion check, not a claim about any real population. Pair it with a one-page fairness policy stub.
Related Lectures
| Lecture | Role |
|---|---|
| Bias | Measurement of gaps that fairness criteria consume |
| Explainability | Local reasons; not a fairness proof |
| Transparency | Publish the chosen criterion and known conflicts |
| Privacy | Whether you may collect slice attributes at all |
| Governance / Responsible AI | Owners, review gates, escalation |
| Compliance | Jurisdiction-specific duties (high level) |
“Fairness means demographic parity, always.” Parity is one template; equal opportunity, equalized odds, calibration, and individual similarity answer different moral questions. Second: maximizing all of them at once—impossibility results say you generally cannot. Third: treating a toy notebook as a published disparity study. Fourth: optimizing a fairness metric on labels that encode the original harm. Fifth: skipping privacy when you start logging race or health attributes “for the dashboard.”
Knowledge Check
- Short Answer: How does fairness differ from bias in this module? Answer: Bias measures gaps; fairness is the chosen goal/policy about which gaps matter.
- True/False: There is one fairness metric that is correct for every AI product. Answer: False.
- Multiple Choice: Equalized odds asks for similar: (a) TPR and FPR across groups, (b) perplexity, (c) token price. Answer: (a).
- Short Answer: What does equal opportunity focus on? Answer: Similar true-positive rates (recall among actual positives) across groups.
- True/False: Demographic parity, equalized odds, and calibration typically all hold together for imperfect predictors with different base rates. Answer: False—they generally conflict.
- Multiple Choice: Individual fairness depends critically on: (a) the similarity metric, (b) BLEU-4, (c) batch size. Answer: (a).
- Short Answer: Why can historical labels undermine equalized odds? Answer: If y encodes an unfair process, matching error rates can reproduce that process.
- True/False: This lecture’s Python table is a real hiring audit. Answer: False—it is explicitly toy/synthetic.
- Multiple Choice: Who should own the chosen fairness criterion? (a) An explicit product/RAI reviewer, (b) nobody, (c) a random intern. Answer: (a).
- Short Answer: Name one sibling lecture that records the chosen criterion for outsiders. Answer: Transparency (model/data cards) or Governance.
Key Takeaways
- Fairness is a signed product goal; bias is the measurement feed.
- Parity, equal opportunity, equalized odds, calibration, and individual similarity are different claims.
- You generally cannot satisfy every group criterion at once; document the trade-off with utility.
- Use toy data to learn the math; never invent fake protected-class studies.
- Next: Explainability — how to talk about why a score happened.
Lab: Same toy frame as Bias. Students write a five-line fairness policy (use case, criterion, non-goals, owner, min-n). Implement the check. Change base rates and show parity vs TPR gap moving opposite directions.
Whiteboard: Three circles—parity, equalized odds, calibration—with “imperfect predictor + different base rates” in the middle. Arrow to Transparency for the written choice.
Recap: Fairness is the criterion you commit to after measuring bias. Pick explicitly, accept impossibility trade-offs, keep examples toy, and continue to Explainability.