A school district moves all homework submission online. Sensible: instant feedback, no lost papers, automatic records.
Except — 12% of the district's families have no reliable home Internet. Some students now do homework in a fast-food parking lot on borrowed Wi-Fi; others take a two-bus trip to the library nightly; some just... stop submitting. Grades in those households drop, not for lack of learning but lack of bandwidth. The innovation worked exactly as designed and still widened a gap — because access to computing is not evenly distributed, and every computing-dependent system inherits that unevenness.
That's the digital divide. Its quieter sibling — bias inside the systems themselves — rides along in this lesson too. Both answer the same exam question: who gets left out, and how did that happen?
The digital divide is the gap in access to computing devices and the Internet — and in the ability to use them — between different groups. The CED's axes of difference:
Key claims to answer with:
The CED's core claim, learn it verbatim-ish: computing innovations can reflect existing human biases, because bias can be embedded at every stage — the data used, the assumptions of the creators, the way the system is tested and deployed. Usually unintentionally.
The entry points, with the standard examples:
| Entry point | How bias gets in | Classic example |
|---|---|---|
| Training/collection data | The system learns from data that under-represents groups or encodes past discrimination | A hiring tool trained on a decade of past hires replicates the old workforce's demographics; facial analysis trained mostly on light-skinned faces errs on dark-skinned faces |
| Creator assumptions | Designers build for users like themselves | Voice assistants that struggle with accents; health sensors calibrated on one skin tone; forms that require a "first/last name" structure some cultures don't use |
| Testing population | The system is validated only on a narrow group | An app tested exclusively on new phones and fast connections fails on older devices — excluding exactly the users on the divide's far side |
| Deployment/feedback loops | Biased outputs generate new biased data | Predictive policing: patrol where past arrests were → more arrests there → data "confirms" the pattern (Lesson 20's scale effects compounding) |
Precision points the exam tests:
The divide and bias interlock: people excluded by the divide are absent from the data, so systems get built and tested without them, so the systems serve them worse, so they benefit less from computing — the divide feeds bias feeds the divide. Passage questions love an innovation that "works great" for its data-rich majority and silently fails a group nobody measured. Your job is to be the analyst who asks: who isn't in this data?
Problem: A telehealth service replaces in-person clinic visits. Name two distinct digital-divide reasons a patient might be unable to use it.
Solution: (1) Infrastructure/cost: no broadband at home or no device capable of video calls. (2) Digital literacy: a patient may own a device but lack the skills (or the accessibility accommodations) to operate the app. Distinct axes — access and ability.
Interpretation: Credited answers name different axes, not two flavors of the same one. "No Internet and slow Internet" is one axis twice.
Problem: A résumé-screening system trained on 15 years of a company's hiring decisions rejects applicants from certain colleges at unusually high rates — colleges whose graduates the company historically rarely hired. Where did the bias enter, and were the developers necessarily biased?
Solution: Entry point: the training data — historical hiring decisions encoded past preferences (perhaps themselves discriminatory), and the system learned to replicate them. The developers need not hold any bias; they inherited it by treating past decisions as ground truth. Fixing it requires auditing/rebalancing the data and validating outcomes across applicant groups — not just "debugging."
Interpretation: The exam's favored answer shape: bias in, bias out, no villain required. When a question asks "how did the system become biased," look for the data answer before any human-intent answer.
Problem: A city launches a parking app and removes coin meters. Which group bears a new burden, and what CED concept applies?
Solution: Residents without smartphones or banking access (skewing elderly and low-income) can no longer park legally without acquiring both. The city converted a universal-access system (coins) into one gated by device + account + connectivity — the digital divide amplified by design: an innovation that presumes access redistributes costs onto those without it.
Interpretation: "The innovation presumes access to X" is the analytical key. Whenever a scenario makes something online-only or app-only, the divide question is being asked.
Problem: A speech-recognition feature performs far worse for speakers with certain accents because its training recordings came overwhelmingly from one region. Which action MOST directly addresses the cause?
(A) Rewriting the code to run faster (B) Expanding the training data to include diverse accents and validating accuracy across speaker groups (C) Adding a disclaimer to the terms of service (D) Encrypting the audio recordings
Solution: (B). The cause is unrepresentative data; the fix is representative data + outcome validation. (A) speeds up the bias; (C) documents it; (D) secures it. None addresses it.
Interpretation: Match the remedy to the entry point. Data problem → data remedy. Design-assumption problem → diverse team/user testing. The distractors will offer remedies for different entry points (or none).
1. (B). Devices + connectivity + skills, between groups. (A)/(C) are vocabulary decoys from other units.
2. (B). Same country, same city even — opposite sides of the access gap. (A) is the between-countries version; the question asked within.
3. (B). The access barrier lands on those least equipped to clear it — the divide's cruelest pattern (and a scenario type the exam reuses: benefits, healthcare, school). The other choices are noise.
4. (C). Data + creator assumptions — the CED's stated entry points. (A) requires intent (wrong), (B)/(D) are jokes (real forms include them; spend zero seconds).
5. (C). Bias without intent, entering through unrepresentative testing/calibration — the precise CED framing. (A) is the intent fallacy; (B) mislabels it a code error (the code did what it was built to do).
6. (D). Output becomes input becomes "confirmation" — the feedback-loop amplifier. (A) mistakes the loop's self-generated data for validation — exactly the trap the loop sets.
7. (D). Infrastructure + affordability + skills: all three divide axes addressed. (B) widens it.
8. (A). Adults-only recordings = unrepresentative training data. (B) blames the excluded users — the pattern to always reject.
9. (A) and (C). Unintentional embedding; deliberate multi-front mitigation. (B) is the intent fallacy again; (D) confuses error-free execution with outcome fairness.
10. (B). Test conditions define who the product works for; high-end-only testing ships the divide inside the app. This is the testing-population entry point, caught in time.
11. (B). Bias hides in representation gaps; the audit starts where the data's blind spots are. The other questions are worthy but answer different concerns.
12. (Model answer.) (i) The training data encodes two decades of the bank's human decisions — if loan officers historically approved fewer loans from certain neighborhoods or demographics (even for reasons since outlawed), the system learns those patterns as "what approval looks like" and replicates them with perfect consistency, no biased line of code required. (ii) Audit outcomes by group: compare approval rates and error rates across demographics/geographies for applicants with similar financial profiles; where disparities appear, rebalance or correct the training data and re-validate — and monitor continuously after deployment. (Any answer pairing a data-inheritance mechanism with a detection/mitigation step at the matching entry point earns full credit.)
Answer letter distribution check: B, B, B, C, C, D, D, A, A+C, B, B — singles: A×1, B×5, C×2, D×2 + multi (A,C). Cumulative through L21 ≈ A 22%, B 32%, C 26%, D 20%.
12 (short response, passage-style). A bank deploys an automated loan-approval system trained on 20 years of its own lending decisions. Describe (i) one specific way bias could be present in this system without any biased code, and (ii) one concrete step the bank could take to detect or reduce it.
Two direct transfers:
Mini practice: in one sentence, name a group your PT idea serves poorly and why. (Model: my recipe-scaling app assumes metric-or-cup units typed in English, so cooks using other languages or unit systems would get wrong results or none.)
1. (B). Devices + connectivity + skills, between groups. (A)/(C) are vocabulary decoys from other units.
2. (B). Same country, same city even — opposite sides of the access gap. (A) is the between-countries version; the question asked within.
3. (B). The access barrier lands on those least equipped to clear it — the divide's cruelest pattern (and a scenario type the exam reuses: benefits, healthcare, school). The other choices are noise.
4. (C). Data + creator assumptions — the CED's stated entry points. (A) requires intent (wrong), (B)/(D) are jokes (real forms include them; spend zero seconds).
5. (C). Bias without intent, entering through unrepresentative testing/calibration — the precise CED framing. (A) is the intent fallacy; (B) mislabels it a code error (the code did what it was built to do).
6. (D). Output becomes input becomes "confirmation" — the feedback-loop amplifier. (A) mistakes the loop's self-generated data for validation — exactly the trap the loop sets.
7. (D). Infrastructure + affordability + skills: all three divide axes addressed. (B) widens it.
8. (A). Adults-only recordings = unrepresentative training data. (B) blames the excluded users — the pattern to always reject.
9. (A) and (C). Unintentional embedding; deliberate multi-front mitigation. (B) is the intent fallacy again; (D) confuses error-free execution with outcome fairness.
10. (B). Test conditions define who the product works for; high-end-only testing ships the divide inside the app. This is the testing-population entry point, caught in time.
11. (B). Bias hides in representation gaps; the audit starts where the data's blind spots are. The other questions are worthy but answer different concerns.
12. (Model answer.) (i) The training data encodes two decades of the bank's human decisions — if loan officers historically approved fewer loans from certain neighborhoods or demographics (even for reasons since outlawed), the system learns those patterns as "what approval looks like" and replicates them with perfect consistency, no biased line of code required. (ii) Audit outcomes by group: compare approval rates and error rates across demographics/geographies for applicants with similar financial profiles; where disparities appear, rebalance or correct the training data and re-validate — and monitor continuously after deployment. (Any answer pairing a data-inheritance mechanism with a detection/mitigation step at the matching entry point earns full credit.)
Answer letter distribution check: B, B, B, C, C, D, D, A, A+C, B, B — singles: A×1, B×5, C×2, D×2 + multi (A,C). Cumulative through L21 ≈ A 22%, B 32%, C 26%, D 20%.
Exam tip: For any bias question, run the three-step: (1) Does the answer require intent? Strike it. (2) Does it call bias a code bug? Strike it. (3) Does it name a data/design/testing entry point and match any proposed fix to that entry point? Credit it. Three strikes of the pencil and the credited answer is usually the one left standing.