D790 Human Centered AI is banner number ITSW 6107 and three competency units in the School of Technology, at the graduate level. The catalog describes it as how AI systems are designed to align with human behavior and values. That is a design course rather than a modeling course. The competency being built is the ability to argue about how a system should behave toward the people it affects, and to turn that argument into design decisions somebody could implement and check.
The people a system acts on are not always its users
The distinction that organizes this subject is between the person operating a system and the person it acts upon. A loan model has an analyst as its user and an applicant as its subject, and the subject has no interface, no choice and usually no visibility. Most of the difficult questions in human centered design concern that second group, and a paper that discusses only the operator has analyzed the easier half.
Appropriate reliance is the design goal worth naming precisely. A system that people trust too much produces automation complacency, where an obviously wrong output is accepted because it came from the machine. A system nobody trusts is ignored and wasted. The design target is calibrated trust, where confidence tracks actual reliability, and the mechanisms that produce it are honest confidence signals, visible uncertainty, and explanations that let a person judge rather than merely feel informed.
Explanation is a design problem rather than a technical one. Different audiences need different things: an operator needs to know what would have changed the outcome, a subject needs to know why this decision and what recourse exists, and a reviewer needs to know how the system behaves across cases. One explanation serving all three serves none, and specifying who each explanation is for is a scored move.
Control and recourse close the design. What can a person override, what happens when they do, how is a contested outcome challenged, and who is accountable for the decision. A system that produces consequential outcomes with no path to challenge them has a design defect, not merely an ethical one, and framing it that way is what makes the argument actionable.
Turning scored aspects into design claims and a budget
Aspects at WGU are judged separately and each needs a 2. In a design and analysis course the mapping runs from aspect to claim, and each claim needs either evidence from a source or a condition drawn from your scenario. Claims supported only by conviction are the ones that come back.
Budget it. Suppose ten scored aspects and a paper of about 2,100 words. Dividing 2,100 across ten leaves 210 per aspect. Sort them. One aspect maps stakeholders and the decision context, 250. Two analyze potential harms and who bears them, at 300 each, or 600. Two cover transparency and explanation design for named audiences, at 280 each, or 560. Two cover human control, oversight and recourse, at 250 each, or 500. One covers evaluation, how you would know the design worked, 200. Two are shorter framing sections at 60 each, or 120. That totals 250 plus 600 plus 560 plus 500 plus 200 plus 120, which is 2,230, so trim one harm section and one transparency section by 60 each, landing near 2,110 with harm analysis and transparency carrying more than half.
Build the stakeholder table before drafting: who is affected, what they can and cannot see, what they can and cannot change, and what a bad outcome costs them. Rows where a person bears a cost without visibility or control are where your harms analysis begins.
On the D790 paper now?
Send the aspects and the system scenario. You get a stakeholder table, a harms analysis structure and transparency decisions written per audience.
Design questions that produce arguable answers
| Question | A weak answer | An answer that scores |
|---|---|---|
| Who is affected without using the system? | Not considered | Named, with what they can see and change stated |
| What does a wrong output cost, and to whom? | Errors are undesirable | Each error type described with who bears it and how badly |
| How does a person know when to doubt it? | The system is accurate | Confidence exposed, and low confidence changing what the interface does |
| What can a person override? | There is human oversight | The specific decisions, the mechanism, and what the override records |
| How does a subject contest an outcome? | They can contact support | A defined route, a responsible role and a stated response time |
| Does it work equally well for everyone? | The model is unbiased | Performance compared across groups, with the comparison shown |
| How would you know the design worked? | User satisfaction | An observable measure of appropriate reliance or successful recourse |
The right column is what a graduate aspect is asking for in each case. The middle column is the version that appears when the paper was written from general principles rather than from the scenario.
Evidence when the subject is values
Arguments about values are still evidence-based arguments, and the strongest material is empirical. Research on automation bias, on how explanations change reliance, and on documented failures of deployed systems gives you findings rather than opinions. Published governance frameworks and risk management guidance give you structures to argue within. Both cite cleanly in APA where your program requires it.
Be precise about fairness, because it has several formal definitions that cannot all hold at once. Equal error rates across groups, equal positive rates and equal calibration are different requirements, and there are impossibility results showing they conflict in ordinary conditions. A paper that names the definition it is using and acknowledges what it gives up is doing graduate work; a paper that uses fair as an adjective is not.
Where you draw on a real deployed system as an example, cite the account you are relying on and describe it accurately rather than dramatically. Where you use your own workplace, anonymize it and present it as one case. And avoid asserting what an AI system understands or wants, because the language of intention smuggles in claims that a technical paper cannot support.
What clears, and what returns
WGU records Competent or Not Competent, with revision available without penalty. Inside a six month flat rate term the cost of a return is calendar days, and in an argument-driven course the difference between a first-pass paper and a returned one is usually whether the claims were grounded in the specific scenario.
Papers that clear name the people affected, describe harms concretely enough to be designed against, specify explanations by audience, define control and recourse as mechanisms rather than intentions, and propose a way to tell whether the design achieved what it claimed. They also acknowledge tension, since transparency, privacy, accuracy and simplicity pull against one another and pretending otherwise weakens the analysis.
Returns come from three habits. Principles restated without application, which produces a paper that would fit any system. Human oversight asserted with no mechanism behind it. And fairness discussed without a definition, which leaves every claim about it unfalsifiable.
Six mistakes that cost D790 students time
- Writing about AI in general. The aspects are about a system in a context. General ethics essays answer a question nobody asked.
- Treating the operator as the only human. The person the system acts on usually carries the greater cost and has the least visibility.
- Oversight as a checkbox. A human in the loop who cannot in practice disagree is not oversight, and saying so is part of the analysis.
- Explanation without an audience. Operator, subject and reviewer need different things, and one explanation for all three helps none of them.
- Fairness as an adjective. Name the definition, state what it costs, and acknowledge the conflict with the other definitions.
- No evaluation. A design proposal with no way to tell whether it worked leaves the closing aspect unanswered.
How we work on this course
D790 support is argument architecture. Send the scored aspects and the system scenario and you get a stakeholder table separating operators from subjects, a harms analysis with costs assigned, transparency decisions written per audience, control and recourse specified as mechanisms, an evaluation proposal, and a model paper in the register WGU graduate evaluators expect. If an objective assessment forms part of your course, readiness work is all we contribute there. Objective assessments are proctored, we get you ready and then step away entirely, and your portal login is never something we ask to see.
Papers about values are easy to write at length and hard to write well. Grounding every claim in your scenario before drafting is what keeps this course inside one six month flat rate term.
Three questions D790 students ask
Is D790 the same course as ITSW 6107?
Is this an ethics course or a technical course?
Do I need to build a model for D790?
Where D790 sits in WGU's programs
The July 2026 catalog places this code in 1 current WGU program. Open a program page for the complete standard path and term positions. The live Degree Plan remains authoritative after transfer credit, substitutions, and mentor planning.
The assessments, one by one
The public catalog does not publish this course's PA/OA identity or task count. WGU Tutors publishes at most one PA manual per course and only from a WGU-controlled public rubric. Until that source exists, PA help begins from the student's real Course of Study and OA support remains preparation only.