Chapter 40 Exercises: AI, Automation, and the Future of Political Analytics

Exercise 40.1 — LLM Political Messaging Experiment (Individual or Pairs, 75 minutes)

This exercise uses an accessible large language model (such as those available through commercial APIs or web interfaces). You will generate political persuasion messages and evaluate them.

Part A: Define a hypothetical voter profile: a 52-year-old woman in a suburban Ohio congressional district, registered independent, with a consumer profile suggesting interest in healthcare costs, small business ownership, and outdoor recreation. She has voted in the last three general elections but skipped the most recent primary.

Part B: Prompt the LLM to write three different persuasion messages from a hypothetical Democratic candidate on the issue of prescription drug pricing. Ask it to write one in a formal policy-focused style, one in a personal narrative style, and one emphasizing economic anxiety.

Part C: Now prompt the same LLM to write three messages on the same issue from a hypothetical Republican candidate, again in three different styles.

Part D: Evaluate each message for: (a) factual accuracy — do any claims need verification?; (b) authenticity — does it sound like a real political message?; (c) potential for harm — could any message be characterized as manipulative?; (d) quality control needs — what would a human reviewer need to check before deployment?

Part E: Write a 400-word memo to a hypothetical campaign manager summarizing what you found and what quality control process you would recommend for deploying LLM-assisted messaging at scale.

Exercise 40.2 — Deepfake Detection and the Liar's Dividend (Group, 60 minutes)

Part A: In groups of three to four, find three to five examples of political deepfakes that have been publicly documented in news coverage. For each, identify: the country/context; the political claim being made; who produced it; how it was circulated; and how it was ultimately identified as synthetic (or whether it was).

Part B: For each example, analyze the liar's dividend effect: could the subject of the deepfake have used it to dispute authentic content? Did they?

Part C: Research one of the following deepfake detection initiatives: the Partnership on AI's synthetic media framework; the Content Authenticity Initiative; or C2PA (Coalition for Content Provenance and Authenticity). Present a 5-minute overview of: what the initiative does; how it works technically; what its limitations are; and whether it would have been effective against the deepfakes you identified in Part A.

Part D: As a group, draft a one-page "platform policy recommendation" on deepfake political content for a hypothetical major social media platform. Your recommendation should address: what content should be prohibited vs. disclosed vs. allowed; how violations should be detected; what the enforcement mechanism should be; and how you would handle edge cases (satire, commentary, historical re-creation).

Exercise 40.3 — Synthetic Respondent Validity Test (Individual, 90 minutes)

This exercise requires access to an LLM API or web interface.

Part A: Identify a political survey question from a publicly released recent poll. Choose a question where you know the actual results for a specific demographic group (e.g., "support for expanding Medicare to cover dental and vision care among registered voters 65 and older in Florida: 67% support").

Part B: Prompt an LLM to simulate responses from 20 individuals in that demographic group, using a prompt like: "You are a 68-year-old retired teacher in Florida who has voted in every election since 1980. On a scale of 1-4 where 1 is strongly oppose and 4 is strongly support, how would you rate your support for expanding Medicare to cover dental and vision care? Explain your reasoning briefly."

Part C: Vary the voter profile significantly across your 20 simulations (different ages, different professions, different locations within Florida, different political leanings within the group). Tabulate the results.

Part D: Compare your synthetic results to the actual poll results. Calculate the difference. Repeat the exercise with a different question — one about a local race or a recently emerged policy issue.

Part E: Write a 300-word methodological assessment: For which types of questions did synthetic respondents perform best? Worst? What would a practitioner need to know about these limitations before relying on synthetic respondent data for political decision-making?

Exercise 40.4 — AI Disclosure Policy Analysis (Individual, 45 minutes)

Review the text of at least two state AI disclosure laws for political advertising that have been enacted as of your current date. (California AB 2839, Michigan HB 6547, and similar laws are good starting points; search for current status.)

For each law, analyze: 1. What content does it require disclosure for? What is the definition of "AI-generated" or equivalent? 2. Who is covered? (Campaigns? PACs? Third-party advertisers? All political advertising?) 3. What form must the disclosure take? (On-screen text? Audio statement? Filed with a regulatory agency?) 4. What enforcement mechanism exists? (Fines? Criminal penalties? Civil liability?) 5. What exemptions or safe harbors exist?

After analyzing both laws, write a 500-word comparative analysis: How do they differ in scope, definition, and enforcement? Which provides stronger protection for voters? What are the most significant gaps in both?

Exercise 40.5 — The Prediction vs. Explanation Trade-off (Pairs, 60 minutes)

You are analytics consultants advising a statewide campaign's voter contact program. The campaign has two models available:

Model A: A gradient-boosted tree model with 87% accuracy on holdout data. It uses 240 features from voter file, consumer data, and behavioral modeling. It cannot generate interpretable explanations of individual predictions.

Model B: A logistic regression model with 79% accuracy on the same holdout data. It uses 12 features, all of which can be clearly explained to a non-technical field director. Its coefficients are interpretable.

With your partner, develop a recommendation for the campaign, addressing:

1. For which specific campaign decisions would you use Model A? Model B? (Consider: setting the overall contact priority list; explaining to field staff why a particular neighborhood is high-priority; making real-time adjustments during the get-out-the-vote period; post-election program evaluation)

2. If Model A is later found to have systematically deprioritized minority-majority precincts (algorithmic bias), how does the lack of interpretability affect your ability to identify and correct this? How does it affect your liability exposure?

3. Draft a brief "model use policy" for the campaign that specifies when each model may be used, what quality control processes are required, and what human review is required before any model output is used in a campaign decision.

Exercise 40.6 — 2030 Scenario Planning (Groups of 4, 90 minutes)

Your group will develop a detailed scenario analysis of political analytics in 2030. Divide into two subgroups:

Subgroup A: Optimistic scenario. AI tools are broadly accessible (even to small campaigns), disclosure requirements are clear and enforced, algorithm auditing is standard practice, and deepfake detection has improved significantly. Describe: (a) what a typical Senate campaign's analytics operation looks like; (b) how polling has changed; (c) what new equity challenges have emerged despite the improved overall landscape; (d) what skills a political analyst hired in 2030 must have.

Subgroup B: Pessimistic scenario. AI tools are concentrated among well-funded campaigns and foreign interference operations, disclosure requirements are weak and inconsistently enforced, the liar's dividend has significantly degraded political accountability, and automated targeting has amplified racial disparities in political representation. Describe: (a) what elections look like in this environment; (b) what the role of human political analysts is when AI optimizes most operational decisions; (c) what institutional responses are possible; (d) what a new analyst should know to work ethically in this landscape.

Reconvene and present both scenarios to the full group. Discussion questions: Which elements of each scenario seem most plausible? What single intervention (regulatory, technical, professional) would do the most to push toward the optimistic scenario? What should political analytics programs be teaching students now to prepare for either scenario?