25%, Individual. Launches Week 3. Due Week 7.

Ethics Case Analysis

A technology was deployed on people who could not easily refuse it. Something went wrong, or something went right, and everyone is pointing at “the algorithm.” Your job is to take the system apart and name the part that decided the outcome.

Assignment

Every one of these systems is a stack of human decisions: who defined the problem, what data was used, what a high score triggers, who reviews the output, and who bears the cost when it fails. “The algorithm did it” is where lazy analysis ends. Your analysis traces the harm (or the success) to a specific decision, with primary-source evidence, and then proposes fixes assigned to the people who could actually make them.

You rehearsed this in the Week 3 lab, where the class dismantled two systems built on the same math, one that triggers investigations (Allegheny) and one that triggers offers of help (LA County). This assignment is the solo performance. The deliverable is a structured document of at most 3 pages (bullets and tables expected) or the stated equivalent for your mode, plus a 5-minute recorded walkthrough.

Download the working template for the default mode: analysis template

Whatever case and mode you choose, four elements are the graded core: run the sociotechnical protocol, work a professional ethics code and its technology standards by name, attribute the outcome to a specific decision with evidence, and propose concrete, role-specific mitigation.

The default code is social work’s, because it is the one with technology standards already written: the NASW Code of Ethics and the NASW Technology Standards. If you practice or plan to practice under a different code, use yours and name it in the submission: the American Public Health Association’s code, the American Counseling Association’s, the ACM Code of Ethics for a computing role, or the code of your own field. Whichever you pick, the same rule applies: five values maximum, each worked against a specific part of the system.

Case selection

Pick from the case bank, Option A or Option B. Off-bank cases are welcome with instructor approval (see FAQ).

Option A: a harm case. Your question is where the harm entered and who decided.

• Michigan MiDAS (unemployment fraud detection, the local case)
• Allegheny Family Screening Tool (child welfare)
• Dutch toeslagenaffaire (child benefits)
• Australia Robodebt (benefits debt recovery)
• SafeRent (tenant screening, 2024 settlement)
• TennCare Connect (Medicaid eligibility)
• UnitedHealth nH Predict (coverage denials)
• COMPAS (criminal justice risk scores)

Option B: a success case, analyzed to the identical standard. Success cases are graded with the same rigor plus the conditions-of-success questions: what made it work, whose participation shaped it, and would it replicate without its champions.

• GetCalFresh / Code for America (benefits access)
• LA County Homelessness Prevention Unit (preventive risk scoring)
• Annie MOORE / GeoMatch (refugee placement matching)
• Trevor Project Crisis Contact Simulator (counselor training)
• Te Hiku Media / Papa Reo (Māori-owned speech AI)
• NVDA screen reader (disability-led open source)
• Be My Eyes / Be My AI (visual assistance)
• M-Pesa (mobile money, contested evidence)

Analysis modes

The mode is only a container for the same analysis. All three have the same four graded elements and the same rubric.

Analysis (default)

The analysis has no external audience and five sections.

1. The system. What it is, who it acts on, and what a high score or output actually triggers in a person’s life.
2. The parts list. The protocol answered for the system: whose problem definition, whose data and labor, what it assumes about users, who bears the risk of failure, what arrangement it locks in.
3. The failure point (or, for Option B, the load-bearing part). The single decision that most separates harm from hope, traced with primary-source evidence.
4. The values diagnosis. Each part examined against your code’s values and its technology standards. Under the NASW default that is dignity, self-determination, social justice, integrity, and competence, worked against the Technology Standards.
5. The mitigation kit. Two or three concrete moves, each assigned to someone who could make it (designer, deploying org, front-line worker, regulator) and tied to the part it fixes.

It is the default because the five sections are the method itself, and the method transfers to any system you meet later.

Consultation note

The note is written to one colleague. Someone at an organization is about to adopt a technology, or was just burned by one, and asks what you think. Write it in the practice register, which here is the situation, background, assessment, and recommendation format clinicians already use: Situation (what they are deciding and why now), Background (the system and its track record, evidenced), Assessment (the protocol as your instrument, your code’s values as the ethical differential), Recommendation (adopt, defer, or refuse, plus the contraindications: the conditions under which you would not adopt, which is your mitigation section). The walkthrough is the verbal hand-off to the requesting practitioner.

Written testimony

Written to a body that decides. A board, city council, or oversight committee is deciding whether to retain, terminate, or repair the technology. You file written testimony (who you are and your standing, the finding, the sociotechnical evidence, the values argument, the specific ask) and deliver it to camera as if at the hearing. The recording is the centerpiece here and absorbs the required walkthrough. This mode suits a case you would rather argue in public than analyze in writing.

Compressed sample analysis

The full brief is up to 3 pages long. This 350-word compression of the Week 3 worked example shows the moves.

1. The system. MiDAS, the Michigan Integrated Data Automated System, was the state unemployment agency’s automated fraud detector. When it spotted a discrepancy between employer records and a claimant’s answers, it did more than flag the file. From 2013 to 2015 it adjudicated: the system itself issued fraud determinations, which triggered penalties of four times the alleged overpayment against people who had just lost a job, collected through garnished wages and seized tax refunds.

2. The parts list. Problem definition: the state’s (catch fraud, cut staffing costs); claimants were never in the room. Data: employer-reported records mismatched against claimant answers, with any discrepancy treated as intent to deceive. Assumption about users: a mismatch means a lie. Risk of failure: borne entirely, and automatically, by the claimant. What it locks in: an agency staffed down on the promise that the machine would decide.

3. The failure point. The failure point is not the matching code. The decision that separates this case from an ordinary screening tool is that auto-adjudication was switched on with no human review of fraud determinations. When the determinations were later reviewed, the error rate was 93%, and roughly 40,000 people had been falsely accused. The same software with a caseworker reviewing every flag is a defensible tool; with the reviewer removed, it accused people automatically.

4. The values diagnosis. Dignity and worth: quadruple penalties treated claimants as presumptive criminals. Self-determination: people were penalized by a process they could not see, question, or answer. Competence (and the Technology Standards): the agency deployed a system whose determinations it could not explain or audit. Social justice: the burden was borne by people already in economic freefall.

5. The mitigation kit. For the deploying agency: no fraud determination issues without individual human review; the flag advises, a person decides. This repairs the failure point directly. For the legislature: mandatory error-rate audits with a threshold that suspends penalties automatically. Michigan was eventually ordered to refund victims; audits move the remedy from after the harm to before it.

Suggested process

1. Pick your case and mode; read the entry-point source in the case bank (1 hour). Log your AI research plan at the Week 3 launch.
2. Gather primary sources (2 to 3 hours). Court rulings, audit reports, agency documents, the underlying investigations. AI can scout here; it is never a source.
3. Run the protocol (90 minutes). Fill every row before you decide anything. The rows you cannot fill tell you what to go find.
4. Find the decision (1 hour of honest thinking). Locate where the harm enters, or which part is decisive. Test it: if that one decision had gone the other way, the outcome should flip too.
5. Run the values diagnosis (1 hour). Work each value from your code against a specific part. If a value produces only a general sentence, cut it and work a different one harder.
6. Build the mitigation kit (1 hour). Every move receives an owner and names a part.
7. Draft into your mode’s structure (2 hours). Tables and bullets are expected; 3 pages is a ceiling.
8. Verification pass (45 minutes). Every claim that came through an AI is checked against a primary source. Write the disclosure line: tool, what you used it for, what it stated wrongly or slanted.
9. Record the walkthrough (1 hour, including the retake you will want).

The whole process requires roughly 10 to 12 hours across the weeks between launch and due date.

Five-minute walkthrough

A good recording is a performance for your mode’s audience, and it is graded on whether a non-technical stakeholder in that role could follow it. Roughly, spend one minute describing the system in plain language, two minutes on the decision and the evidence for it, one minute on the values stakes, and one minute on what to do and who does it. Do not read your document aloud. Talk to the person who has to trust or distrust this machine tomorrow morning.

Rubric (100 points, scaled to 25%)

The assignment is graded in two pieces. Fifteen points are awarded in Week 4, on the case you picked and the plan for the evidence, so that a case that cannot be sourced is caught while there is still time to change it.

Case and research plan, due Week 4 (15 points)

Analysis, due Week 7 (85 points)

Frequent errors

• Blaming “the algorithm.” The 25-point line item asks for a decision: a design choice, a data choice, a trigger, a deployment choice, an oversight choice. “The AI was biased” names nothing anyone can fix.
• Decorative ethics. Five values pasted as a closing list score as decoration. They earn their 20 points only when each one is worked against a specific part of the system.
• Unverified AI claims. One confident, wrong, unchecked fact in your evidence section costs you there and undermines everything around it.
• The plant. Some case packets contain one plausible but false “fact” that AI research assistants will cheerfully repeat. Catching it against a primary source earns verification points, and repeating it costs them.
• Mitigation without an owner. “There should be more oversight” leaves the row empty, where “the county board requires an annual disparate-impact audit before contract renewal” is a move.
• Option B as press release. A success case retold in the vendor’s voice misses the assignment. You are asked what made it work and whether it would survive its champions leaving.

FAQ

Can I analyze a case that is not in the bank? Yes, with approval. Pitch it at the Week 3 launch: name the system, the population it acts on, and two primary sources you already have. If the sources exist, the answer is almost always yes.

Can I pick MiDAS? You can, but the bar is higher: the compressed sample above and the Week 3 seminar walkthrough already do the obvious version. You would need a fresh angle, for example the seven-year accountability fight or the design of the remedy.

What does “3 pages or the stated equivalent” mean? The analysis and the consultation note are documents: 3 pages maximum, structured, tables and bullets welcome. Written testimony is a filing of about 2 pages plus the recorded delivery, which is the centerpiece there.

What counts as a primary source? Court rulings and filings, government audits, agency documents, the original investigative reporting (with documents), peer-reviewed evaluations, and first-party statements from the organizations involved. A chatbot’s answer is not a source. A blog post summarizing a news story that summarizes an audit is two steps too far; go to the audit.

How does AI disclosure work? A short disclosure line on the submission, in the format the class charters in Week 5: which tool, what you used it for, and what it stated wrongly or slanted. The field for what it stated wrongly is required; leaving it blank counts as not having checked.

Do I have to use the NASW code? No. It is the default because it is the code with technology standards already attached, and the Week 2 material sets it up. If you work under another professional code, use it and name it. Picking one code and working it is not optional; a homemade list of personal values does not satisfy the 20-point criterion.

Is a success case easier? No. Option B has the same protocol, the same evidence standard, and the extra conditions-of-success questions. Praising a system is easy; explaining why it endures is the hard, gradeable part.