MPA 2479 · Applied AI for Public Policy

Master of Public Affairs · Brown University · Fall 2026

Course plan authors: Han Zhang and Codex (OpenAI)

Thursdays, 9:00–11:30 am · 280 Brook Street, Room 110 · 12 class meetings

Office hours: Thursdays, 2:00–4:00 pm · Watson 242

The course at a glance

	W1	W2	W3	W4	W5	W6	W7	W8	W9	W10	W11	W12
	Learning the AI agent			Using AI for evidence				Government AI	Project design	AI governance	Project review and presentation
AI agent what students learn to do	What an AI agent is	Git and version control	Configuring the agent	AI for data collection	AI for data coding	AI for visualization and publishing	Algorithmic bias	Auditing a federal AI system	Designing your final project	AI governance debates	Code review and replication	Final presentations
Policy analysis Bardach & Patashnik eightfold path	Introduction to the eightfold path	Your project: problem definition and the eight steps	Problem definition and initial alternatives	Evidence and the policy question	Measures and evaluation criteria	Findings, alternatives, and communication	Errors, criteria, and trade-offs	Government AI: criteria and alternatives	Alternatives, criteria, and possible outcomes	AI governance debates	Replication and revision	The eightfold path, revisited
	Introducing the framework	Discussing actual project work						Government AI policy	Comparing alternatives	AI governance	Reviewing and communicating findings

The policy row shows discussion focuses. Return to any of the eight steps as your project develops.

Agentic AI tools, policy analysis, and the policy of AI

Learn agentic AI tools through actual tasks. Follow a demonstration and practice the same method using the example data or suitable data from your own project. Weeks 1–3 use life expectancy data from Our World in Data. Weeks 4–5 use Federal Register rules and their abstracts for collection and text coding.

Discuss what agentic AI can and cannot do for policy analysis, using the eightfold path of Bardach and Patashnik. Week 1 introduces the eight steps in class. From Week 2, each group applies the steps to its own project, starting with a basic problem definition. Each week, groups bring an actual result or an unresolved question, and report where AI helped and where it did not. The discussion draws on your experience of using the agentic AI tools on your project, recorded in the weekly memos, and on a reading in the weeks that have one. Groups return to earlier steps as the question and the evidence develop.

Study the policy of AI. Weeks 7, 8, and 10 treat AI systems as the object of policy. Readings and cases cover algorithmic bias, AI systems in government, and the alternatives for governing AI.

Define the problem
Assemble evidence
Construct alternatives
Select criteria
Project outcomes
Confront trade-offs
Decide
Tell the story

These are the recurring questions for every project. Groups may work on several steps at once and revisit them through the semester. Alternatives are the policy options a decision-maker could choose.

A typical 150-minute class

Activity	Minutes
Learning agentic AI tools	45
Hands-on exercises with the tools	45
Discussion of readings, projects, and cases	45

This is a guide, not a rule. Early weeks spend more time learning the tools. Later weeks spend more time in discussion.

Assignments and grading

Group project: 50%. Groups of two or three develop a policy product for a named audience, supported by a GitHub repository. Three checkpoints are worth 5% each, the presentation 10%, and the final product and GitHub repository 25%.

The policy product. It takes one of two forms.

Form	What you submit
Decision memo	PDF, three to five pages, plus a technical appendix
Web page or app	A public web page, a dashboard, or a small application. Any form the audience can open in a browser.

Whichever form, the product follows the eight steps of policy analysis. Each step appears in the product, in the order below, however briefly:

Define the problem, and name the audience.
Assemble the evidence: where the data came from, how you collected them, and what they show.
Construct the alternatives.
Select the criteria.
Project the outcomes of each alternative.
Confront the trade-offs.
Decide, and say what the analysis cannot answer.
Tell the story to the audience you named.

Add one short section on the agent: which steps you gave to it, one concrete example of its contribution or its limit, what you checked, and what you did next.

The GitHub repository. It contains the guidance another analyst needs to replicate the numbers and results in your product, and the code itself if you built a web page or an app.

Weekly memos: 30%. Write 1–2 observations about your project or a classroom case: which policy-analysis step you attempted, how AI helped or fell short, and what you would do next. Aim for 150–200 words total. Describe what you actually tried and observed in your own words. Do not use AI to write the memo. Begin in class and submit to Canvas Discussions on the dates below. Grading rewards specific examples and thoughtful interpretation.

Classwork and participation: 20%. Complete the hands-on exercises, policy discussions, peer review, and replication in class. Grading considers your contribution and ability to explain the work.

Calendar: no class Oct 8 or Nov 26. Final product and GitHub repository due Dec 17.

Week	Topics and agentic AI tools	Hands-on	Project and policy discussion	Assignments
Week 1Sept 10	What an AI agent is Set up Codex or Claude Code and run a first policy-data task on your own computer. Follow an agent as it reads a file, writes and runs code, analyzes the data, and makes a chart. Learn how to direct a task and review a proposed plan. Either tool works. Codex is recommended because students can use it free. Each tool comes as a terminal program and as an app, and either form works. The Codex setup guide and the course setup guide cover both.	Download the Our World in Data life expectancy data (CSV). Plot life expectancy from 2000–2023 for the United States, Canada, the United Kingdom, Germany, and Japan. Compare the change since 2000 and the gap between the United States and the other countries. Check one plotted value against the source. Ask the agent to change the comparison countries or time window, and inspect the revised script and chart.	AI and the eightfold path Introduce the eight steps using one policy question: how could public policy improve life expectancy in the United States? Read in class: Margetts & Dorobantu, Rethink Government with AI (Nature, 2019), 3 pages. For each use of AI in the article, which step is it? Which steps have no example? Use today's analysis and chart as examples of where AI may help. For each of the eight steps, vote Useful, Not useful, or Unsure and discuss your reasons. Record the class tally and a few reasons. Repeat the vote on the same question in Week 12 and compare judgments using the semester's work. Preview Week 2: discuss your own project, define the problem, try applying the steps, and begin the first memo.
	Learning goalSet up Codex or Claude Code and run a first policy-data task on your own computer. Analyze life expectancy data, make a chart, and check the results. Form an initial judgment about AI's usefulness across the eightfold path.
Week 2Sept 17	Git and version control Use Git to record changes, return to a working version, and share work. Direct a multi-step task: explore the files, plan the work, execute it, and commit the result. Review what belongs in a shared GitHub repository: scripts, instructions, and documentation. Keep credentials out of Git.	Continue the life expectancy analysis. Change the comparison countries or years and commit the script and explanation to your GitHub repository. Inspect the difference between versions and restore a working script. Work in pairs in a shared practice GitHub repository. Each person contributes a change; resolve a conflict and check the script.	Your project: problem definition and the eight steps Write a basic problem definition for your own project: what is happening, to whom, where, and who could act on it? Try applying the eight steps. What evidence might you need, what policy alternatives could be considered, and which questions are still open? Try using AI on one part of this initial analysis. What did it add, miss, or leave for you to decide? Discuss your idea with a classmate, revise the problem statement, and record 1–2 observations for your first memo.	Weekly memo · Sept 24 before class.
	Learning goalUse Git to manage an analysis, define an initial problem for your own project, and reflect on a first attempt to apply the eight steps with AI.
Week 3Sept 24	Configuring the agent Write project instructions for the question, data conventions, output requirements, and permitted actions. Introduce instructions files, memory, and Skills. Review the documentation for your agentic AI tools. Compare outputs with and without explicit instructions. Check the choices that were followed and those that still need attention. Rerun the saved script with the same inputs; record any dependencies needed to reproduce its output.	Use the life expectancy analysis to write instructions for country selection, year range, units, source citations, and output format. Run the task, compare outputs, and revise one unclear requirement. Write a short Skill for checking the saved workflow. Form a project group. Refine the problem discussed in Week 2 and identify a data source, audience, and intended product.	Problem definition and initial alternatives How does the question change when you specify an audience, population, place, or time period? What evidence would you need? What policy alternatives could the decision-maker consider? Which written instructions helped with the actual task? What remained a choice for your group? Refer to Bardach on problem definition as needed. Use the instructions, memory, and Skills documentation linked on the week page.	Checkpoint 1: question and data · 5% Group GitHub repository and README: policy question, audience, proposed data source, and intended product. Due Oct 1 before class. Weekly memo · Oct 1 before class.
	Learning goalWrite instructions grounded in a task you have already tried, and leave with a provisional question and a feasible route to project data.
Week 4Oct 1	AI for data collection Use an API or collect a table from a public source. Save the raw response and record its URL, query parameters, and retrieval date. Extend a working collection script to additional places or years. Use subagents for independent subtasks when useful, then check the combined output. Document dependencies and the commands needed to rerun the collection.	Use the Federal Register API to collect 100 rules and proposed rules published in 2025 for a chosen agency. Save document IDs, dates, agencies, titles, abstracts, and source links. Check several records against their published pages. Change the agency or date range and rerun the collection. Exchange GitHub repositories and rerun a classmate’s collection using the README. The abstracts become the text sample for Week 5. Adapt the collection workflow to your group’s data source for Checkpoint 2.	Evidence and the policy question What does the source record, and who or what is absent? Does the available evidence fit your question? How might it change the problem or the alternatives worth considering? Where did AI help you find or collect evidence, and where did you need additional work? Gebru et al., “Datasheets for Datasets” (2021). Focus on what to document about the data you collect.	Checkpoint 2: data collection · 5% Project collection script, source records, and a README that lets another person rerun the workflow. Due Oct 15 before class. Weekly memo · Oct 15 before class.
	Learning goalCollect a documented dataset and explain how its coverage supports or changes the project question.
Oct 8: no class. The Week 4 checkpoint and short memo are due Oct 15.
Week 5Oct 15	AI for data coding and model choice Define a codebook for turning text into variables. Pilot the codebook on a small sample; compare human and model codes and revise unclear categories. Freeze the codebook before evaluating it on a separate human-coded sample. Use a script and a model API for batch coding. Compare errors, a simple baseline, cost, and data-use requirements when choosing a model.	Use the Federal Register abstracts collected in Week 4. In pairs, read 10 abstracts and define categories for the policy action each document describes. Pilot the coding with AI, inspect disagreements, and revise the codebook. Freeze the rules, then hand-code 20 different abstracts before seeing their model labels. Have the agent write and run a script to code those 20 abstracts through a model API. Compare the model with the human codes and a simple keyword baseline. Inspect the confusion matrix and explain two disagreements. Try a larger batch or a second model as an extension.	Measures and evaluation criteria What does the coded variable measure, and how would it help answer your question or compare alternatives? Which errors matter for that use? Does your validation support using these codes? What did AI make easier, and what did defining and checking the measure still require? Gilardi, Alizadeh & Kubli, “ChatGPT outperforms crowd workers for text-annotation tasks” (2023). Focus on how annotation performance is evaluated.	Weekly memo · Oct 22 before class.
	Learning goalConstruct a measure, evaluate it on held-out human codes, and explain whether it is useful for the intended policy analysis.
Week 6Oct 22	AI for visualization and publishing Analyze the collected data and choose a figure that answers a specific question. Recognize hallucination in AI-generated explanations. Check figures, captions, and source claims against the evidence. Use the agent to build an HTML results page and publish it.	Build a results page from your project data, the life expectancy series, or the Federal Register records. Choose a policy question, make one figure, and write a caption stating the finding and its audience. Check the figure against the data. Trace two claims in an AI-generated explanation to the script output or an original source. Revise unsupported or inaccurate statements before publishing. Publish the page from your GitHub repository and exchange feedback. Develop it into the first version of your group’s policy product.	Findings, alternatives, and communication What does the result establish? What remains unresolved? Which policy alternatives or evaluation criteria does it inform? What additional evidence would you need to project their outcomes? Did AI help you analyze or communicate the result? Show a concrete example.	Checkpoint 3: preliminary results · 5% A script-produced figure, the finding it supports, and a first version of the group’s policy product. Due Oct 29 before class. Weekly memo · Oct 29 before class.
	Learning goalProduce a checked finding for a specific reader and explain its relevance to the policy choices being considered.
Week 7Oct 29	Algorithmic bias Compare overall and subgroup errors in a suitable classification example. Use the COMPAS debate to examine how different fairness criteria lead to different assessments.	Use ProPublica’s COMPAS data and analysis (CSV) to compare false positive rates and predictive precision across groups, following the published sample filters. Compare possible uses of the score and explain which errors matter for each policy option.	Errors, criteria, and trade-offs How would an error change your finding or recommendation? What criterion is each side of the fairness debate using? Which matters for the decision being made? Which changes would strengthen your own analysis, and how much has AI helped with checking it? Angwin et al., “Machine Bias” (ProPublica, 2016). Feller, Pierson, Corbett-Davies & Goel, the response to ProPublica (Washington Post, 2016).	Weekly memo · Nov 5 before class.
	Learning goalUse actual checks and contrasting criteria to assess what the analysis supports and what should change.
Week 8Nov 5	Auditing a federal AI system Apply the collection and checking skills to a government AI case. Read OMB’s 2025 AI use-case inventory alongside the April 2025 M-25-21 requirements. Assess documented practices and identify information still needed to judge a system.	Download OMB’s 2025 AI use-case inventory (CSV) and select one high-impact system. Build a scorecard comparing its reported practices with the assigned M-25-21 criteria. Cite the evidence and identify a documented gap or a question for further investigation. Discuss the scorecard in class and compare possible responses. Form Week 10 groups and choose a governance debate.	Government AI: criteria and alternatives What can the published inventory establish about this system? What would require further investigation? What alternatives could the agency consider, and which criteria should govern that choice? Where did the agent help collect or assess the evidence? OMB M-25-21 (April 2025), high-impact AI definition and minimum risk-management practices. Engstrom, Ho, Sharkey & Cuéllar, Government by Algorithm (2020), pp. 6–20.	Weekly memo · Nov 19 before class.
	Learning goalAssess a documented government AI case and explain which judgments the available evidence permits.
Week 9Nov 12	Designing your final project Use the session for project support and individual questions. Review the work already completed and choose the most useful next analysis or revision.	Continue the existing project: resolve a data problem, improve an analysis, compare an alternative, or revise the product. Update the instructions and README to reflect decisions made since Week 3. Bring one unresolved issue to discuss with the instructor.	Project progress across the eight steps Which parts of the analysis are ready to use? Which step needs more work? Can you compare policy alternatives or form a recommendation with the evidence you have? Choose a feasible next task and decide how AI might assist.
	Learning goalMake a feasible revision to an existing project or identify the work needed next.
Week 10Nov 19	AI governance Use Bradford’s comparative framework to introduce the governance discussion. Groups investigate one of six areas: frontier-model risk; jurisdiction and regimes; geopolitics and chips; power and concentration; the information ecosystem; or labor and the economy. Use the agent to locate and organize sources, and check the claims used in the presentation.	Each group presents for 8–10 minutes: a concrete policy dispute, the actors, a sourced case, and one core reading. In class, build a comparison table for at least two policy alternatives, using stated criteria and evidence. Recommend an option to a named decision-maker and discuss the trade-off most relevant to that choice.	Alternatives, trade-offs, and decisions What does each alternative improve, and what costs or risks does it introduce? Would a different criterion change the recommendation? How did AI contribute to the group’s evidence, alternatives, or comparison? Bradford, Digital Empires (2023), Introduction. One core reading selected by your group for its debate. Bring the citation and source. The AI Index policy chapter and the existing reading pool are references.	Weekly memo · Dec 3 before class.
	Learning goalCompare policy alternatives and defend a recommendation while explaining how AI contributed to the analysis.
Nov 26: Thanksgiving recess. Next class Dec 3.
Week 11Dec 3	Code review and replication Read a project through its README, output-producing scripts, and data documentation. Use the session for replication, feedback, and revision.	Exchange group GitHub repositories and run a classmate’s project using the documented commands. Record what ran, what failed, and what information you had to infer. Share the feedback and work on corrections in class. Review the group’s finding and its relevance to the policy options. Explain which claims the evidence supports and suggest a useful revision. Use the review to revise your group project for the presentation and final submission.	Replication and revision Can another analyst reproduce the result and understand the choices behind it? What did the agent help diagnose or repair? Which part of the analysis or recommendation still needs attention?	Weekly memo · Dec 10 before class.
	Learning goalUse a classmate’s rerun to improve both the project and its explanation.
Week 12Dec 10	Final presentations Presentations and questions fill the session, with a break and a closing discussion. Repeat the Week 1 vote on AI's usefulness at each policy-analysis step. Compare the class tallies and reasons using examples from the semester's work.	Present the question, evidence, findings, and chosen policy product. Include one concrete example of AI’s contribution or limitation, what you checked, and what you did next. Record the questions and revisions to address before the final submission.	The eightfold path, revisited Return to the Week 1 question: how could public policy improve life expectancy in the United States? Vote Useful, Not useful, or Unsure for each step before viewing the original tally. Compare the two votes. Explain one judgment you kept or changed using an example from your project, class exercises, or memos. What can your final product support for its intended reader?	Group presentation · 10%Dec 10, in class.Final product and GitHub repository · 25% Policy product, scripts, source information, and instructions for reproducing the results. Due Dec 17.
	Learning goalCommunicate a defensible policy analysis and assess AI’s role using the work completed during the semester.