Course Planning: DTPA Use Case Matrix

This document acts as the central syllabus planning ledger for CMSE 101. It inventories core AI use cases distributed across different socio-technical domains. Each case study is mapped through the Data-Tools-Practices-Actions (DTPA) framework to provide a comprehensive audit of how AI systems are designed, deployed, and experienced in real-world contexts.

Perspective Markers Key

Marker	Curricular Meaning
🚀 HYPE	Unpacks corporate marketing narratives of “infinite growth” vs. material realities.
🏛️ STATE	Highlights deployments prioritizing institutional surveillance, carceral tracking, or control.
⬛ BOX	Identifies black-box opacity driven by trade secrets or mathematical unexplainability.
🌳 SYSTEM	Evaluates long-tail structural externalities (labor erosion, planetary carbon costs, systemic inequity).

Domain Overview & Use Case Links

🎓 Domain 1: Education

1.1 Automated Essay Scoring — Surface-feature regression models vs. dialect bias and Perelman’s BABEL generator text.
1.2 AI Tutoring Systems — Conversational Socratic LLMs (Khanmigo) and the risks of equity substitution in underfunded districts.
1.3 AI-Powered Exam Proctoring — Computer vision behavioral classifiers turning students’ homes into spaces of biometric surveillance.
1.4 Predictive Student Retention — Early warning systems reifying historical structural socioeconomic vulnerabilities as individual risk scores.

🏥 Domain 2: Healthcare

2.1 Medical Image Diagnosis — Deep learning radiology diagnostics vs. training set data gaps for rare conditions and diverse skin tones.
2.2 Health Insurance Risk Scoring — Dissecting the landmark Obermeyer et al. (2019) study where billing cost proxies automated systemic medical racism.
2.3 Mental Health Chatbots — Scripted CBT wellness apps deployed to bridge systemic shortages in human therapeutic infrastructure.
2.4 Sepsis Prediction — Real-time hospital clinical alerting tools navigating the high-stakes friction between alarm fatigue and false negatives.

⚖️ Domain 3: Criminal Justice & Surveillance

3.1 Predictive Policing — Spatial-temporal ETAS hot-spot algorithms trapping communities in self-fulfilling carceral feedback loops.
3.2 Facial Recognition in Law Enforcement — Algorithmic misidentification pipelines causing wrongful arrests of Black citizens (e.g., Robert Williams case).
3.3 Risk Assessment at Sentencing (COMPAS) — Proprietary recidivism scoring models illustrating the mathematical incompatibility of fairness metrics.
3.4 Mass Surveillance & Smart Cities — The modern urban surveillance stack integrating acoustics, Stingrays, and multi-vendor sensor data streams.
3.5 Automated License Plate Readers (ALPRs) — Massive vehicular tracking networks monitoring public spaces and drawing cross-jurisdictional geographic perimeters.

🎨 Domain 4: Art & Creative Industries

4.1 Generative Image Models — Diffusion networks trained on non-consensual web scrapes vs. intellectual property and artist labor models.
4.2 Actor Likenesses & SAG Strike — Generative digital avatars, studio contract negotiations, and the ownership over a performer’s physical identity.
4.3 AI in Legal Knowledge Work — The Mata v. Avianca case exposing next-token loss function confabulations in professional research pipelines.

🌍 Domain 5: Environment & Climate

5.1 Climate Modeling and Weather Prediction — Graph Neural Networks emulating fluid physics vs. regional data starvation in the Global South.
5.2 Carbon Cost of AI — Tracking water footprints, data center grid infrastructure strain, and the material cost of model training cycles.

💼 Domain 6: Labor & Workplace Analytics

6.1 AI Resume Screening & Hiring — Automated candidate sourcing pipelines encoding historical corporate gender and racial demographic baselines.
6.2 Algorithmic Management & Gig Economy — On-demand application dispatchers optimizing continuous piece-rate wages and platform surveillance.
6.3 Sub-Second Productivity Surveillance — Comprehensive keystroke tracking, desk cameras, and mouse metrics intensifying white-collar workplace acceleration.

7.1 Recommendation Algorithms — Engagement optimization logic maximizing screen time, radicalization paths, and societal attention capture.
7.2 Deepfakes & Non-Consensual Media — Generative face-swapping pipelines weaponized disproportionately for harassment, blackmail, and political disruption.

🏠 Domain 8: Housing & Real Estate

8.1 Algorithmic Rent Pricing — Yield management platforms (RealPage) processing non-public lease data to coordinate synthetic market-wide rent inflation.

🌐 Domain 9: Global Supply Chains & Identity

9.1 Ghost Work & Invisible Labor — The distributed international underclass executing traumatizing content moderation and reinforcement learning labels.
9.2 Biometric ID Systems (Aadhaar) — High-stakes identity tracking infrastructure conditioning state survival resources on biometric validation performance.

💳 Domain 10: Finance & Commerce

10.1 Alternative Data Credit Scoring — Ingesting utility histories and digital exhaust data to construct algorithmic proxies for creditworthiness.

🪖 Domain 11: Geopolitics & Warfare

11.1 Lethal Autonomous Weapons Systems (LAWS) — Vision-guided drone swarms and automated targeting matching pipelines completely decoupling human decision-making from lethal kinetic actions.

Cross-Cutting Themes

Each use-case is mapped to one or more recurring themes that surface across domains. These themes provide the analytical scaffolding for comparing cases and structuring class discussion.

Theme 1: Feedback Loops

Algorithmic outputs alter the world in ways that generate the next round of training data, producing self-fulfilling predictions.

1.4 Predictive Student Retention — “At-risk” labels alter institutional behavior, depressing outcomes and confirming the model.
2.2 Health Insurance Risk Scoring — Denying preventative care drives erratic, emergency-only cost data back into the system.
3.1 Predictive Policing — The quintessential loop where output dictates the creation of its own future training data.
3.3 Risk Assessment at Sentencing (COMPAS) — High risk scores destabilize lives, manufacturing the future arrests that validate the model.
7.1 Recommendation Algorithms — Polarizing engagement shifts user baselines, generating more toxic interaction data.

Theme 1 (variant): The Illusion of Accuracy

Headline accuracy numbers conceal high-cost failure modes for the people on the receiving end.

3.5 Automated License Plate Readers (ALPRs) — OCR misreads (O vs. Q) trigger felony-style stops of innocent drivers.
11.1 Lethal Autonomous Weapons Systems (LAWS) — A “94% accurate” classifier means a 6% error rate that manifests as civilian deaths.

Theme 2: Proxy Variables

Easily measurable features stand in for unmeasurable social phenomena, encoding structural inequities as individual traits.

1.1 Automated Essay Scoring — Lexical complexity and sentence length proxy for “critical thinking.”
1.4 Predictive Student Retention — Historical financial strain proxies for personal academic aptitude.
2.2 Health Insurance Risk Scoring — Billing cost stands in for physical health need.
2.4 Sepsis Prediction — Lab-sample logging timestamps proxy for biological decay.
6.1 AI Resume Screening & Hiring — Language choices and college names proxy for gender.
8.1 Algorithmic Rent Pricing — Vacancy rates proxy for a community’s maximum economic pain threshold.
10.1 Alternative Data Credit Scoring — Phone charging logs and social graphs proxy for class and race.

Theme 3: The Benchmark Illusion

Strong performance on curated benchmarks masks brittle behavior in messy real-world deployment.

1.1 Automated Essay Scoring — High machine-human correlation ($r \geq 0.85$) rewards surface conformity, not depth.
2.1 Medical Image Diagnosis — AUC scores drop sharply across scanner hardware and patient diversity.
3.2 Facial Recognition in Law Enforcement — Lab accuracy collapses on low-resolution real-world camera footage.
3.3 Risk Assessment at Sentencing (COMPAS) — Aggregate AUC conceals unequal distribution of false-positive costs onto Black defendants.
4.3 AI in Legal Knowledge Work — 90th-percentile bar exam scores; catastrophic failure in real legal research.
5.1 Climate Modeling and Weather Prediction — GraphCast’s global metrics conceal severe accuracy drops over uninstrumented regions.

Subjects cannot meaningfully refuse the system; opting out means losing the underlying good (school, healthcare, housing, citizenship, public space).

1.3 AI-Powered Exam Proctoring — Consent to bedroom surveillance or fail the course.
2.3 Mental Health Chatbots — Low-income users default to bots because human therapy is unaffordable.
2.4 Sepsis Prediction — Patients unaware their triage is shaped by a proprietary score.
3.1 Predictive Policing — Targeted neighborhoods get heightened surveillance with no recourse.
3.4 Mass Surveillance & Smart Cities — Residents cannot opt out of municipal microphone and camera arrays.
3.5 Automated License Plate Readers (ALPRs) — Cannot opt out without abandoning public roads entirely.
4.2 Actor Likenesses & SAG Strike — Opaque contract terms transfer permanent ownership of a performer’s likeness.
6.2 Algorithmic Management & Gig Economy — Workers must surrender behavioral data to earn a living.
7.2 Deepfakes & Non-Consensual Media — Targets cannot protect their faces and voices from ingestion.
8.1 Algorithmic Rent Pricing — Tenants excluded from the data-pooling scheme that sets their rent.
9.2 Biometric ID Systems (Aadhaar) — Refusing biometric enrollment means losing food security and healthcare.
10.1 Alternative Data Credit Scoring — Low-income applicants must surrender digital privacy for emergency capital.
11.1 Lethal Autonomous Weapons Systems (LAWS) — Civilians in automated combat zones have zero recourse.

Theme 5: Automation Bias

Human decision-makers defer to the model’s output, treating algorithmic flags as authoritative and transferring responsibility onto an opaque system.

1.2 AI Tutoring Systems — Instructors accept dashboard “mastery scores” as flawless.
1.3 AI-Powered Exam Proctoring — Instructors defer to cheating scores instead of investigating.
2.1 Medical Image Diagnosis — Over-extended clinicians use algorithmic flags as a shortcut.
3.2 Facial Recognition in Law Enforcement — Officers defer to the match instead of investigating.
4.3 AI in Legal Knowledge Work — Attorneys trust LLM confidence over opposing counsel’s warnings.
6.1 AI Resume Screening & Hiring — Recruiters uncritically accept “top talent” sorting.
7.1 Recommendation Algorithms — Users absorb algorithmically curated content as personally relevant.

Theme 6: The Democratization / Displacement Tension

Tools framed as expanding access simultaneously dismantle the human labor and institutional infrastructure they claim to extend.

1.2 AI Tutoring Systems — 24/7 homework support vs. displacement of human educators in underfunded districts.
2.3 Mental Health Chatbots — “Democratized access” rhetoric used to justify eliminating skilled human care labor.
4.1 Generative Image Models — Lowers barriers for some while destroying the livelihoods of the artists whose data trained it.
4.2 Actor Likenesses & SAG Strike — Shifts economic returns from working actors to studio technology platforms.
6.3 Sub-Second Productivity Surveillance — Transforms remaining human roles into monitored extensions of the machinery.
9.2 Biometric ID Systems (Aadhaar) — Streamlines state administration while displacing safety nets for the poorest.

Theme 7: Invisible Labor

Automated systems depend on hidden, low-wage human workforces whose presence is deliberately obscured to preserve the “AI” narrative.

3.4 Mass Surveillance & Smart Cities — Acoustic flags routed to hidden review centers for sub-second human override.
6.2 Algorithmic Management & Gig Economy — Workers absorb depreciation costs while isolated from institutional support.
9.1 Ghost Work & Invisible Labor — The primary case: tech infrastructures masking dependence on human workforces.

Theme 8: The Carbon-Justice Contradiction

Commercial expansion of automated products drives environmental degradation while marketing itself as a sustainability solution.

5.2 Carbon Cost of AI — Water footprints, grid strain, and training-cycle costs of the AI build-out.