AI for Decision Makers
Unit 8: Political Systems
AI touches political life in two ways. Indirectly, it shapes the information environment through ranked feeds, content moderation, and cheap generated content. Directly, it shapes how the state treats citizens, through automated eligibility checks, fraud detection, and debt recovery. Public debate concentrates on the first; the clearest documented harms are in the second. This unit treats both with the same evidentiary standard.
The Institutional Landscape
| Domain | AI application | Democratic stake |
|---|---|---|
| News & information | Engagement-ranked feeds and recommendations | What citizens see, and what gets rewarded for being produced |
| Content moderation | Classifiers that remove, demote, or label content | Private enforcement of speech rules at global scale |
| Campaigns | Targeted and generated political messaging | Persuasion, transparency, and the cost of producing propaganda |
| Public administration | Automated eligibility, fraud scoring, debt recovery | Due process and the legality of state decisions about individuals |
Decision: Content Moderation
Large platforms handle volumes no human workforce could review, so automated classification is unavoidable. The governance question is how errors are distributed and corrected. Every threshold trades wrongful removals against harmful content left up (Unit 3). The costs of each error fall on different people and differ by language, culture, and context.
Content Moderation Through the Five Stages
| Stage | Traditional | AI-Mediated | Risk |
|---|---|---|---|
| 🔍Framing | Community standards written by policy teams, interpreted by reviewers | Standards operationalized into categories a classifier can learn | Context-dependent categories (satire, reporting on violence, reclaimed slurs) resist operationalization |
| 📊Prediction | Human reviewers, often under heavy time pressure, with inconsistent judgments | Classifiers score the probability that content violates each policy | Accuracy is lower in languages and dialects with little training data |
| ⚖️Decision rule | Escalation rules and reviewer discretion | Thresholds for automatic removal, demotion, or human review | Uniform thresholds distribute errors unevenly across communities |
| 🎯Action | Remove, keep, label, or escalate | Automated removal or reduced distribution at scale | Fast errors reach millions before correction; demotion is invisible to the user affected |
| 🔄Feedback & adaptation | Appeals and policy revisions | Appeals and reports become training data; adversaries probe the classifier | Coordinated reporting and evasion tactics adapt to the system (interdependence) |
Moderation is also a strategic interaction among platforms, users, and regulators. Stricter rules on one platform push some users and creators to others. Regulation in one jurisdiction changes global defaults or shifts activity elsewhere. The EU's Digital Services Act now requires very large platforms to assess and mitigate systemic risks, including risks to civic discourse and electoral processes. It also requires independent audits and access to platform data for vetted researchers. That last provision matters because so much of the debate below has been conducted without data.
Case Study: Engagement-Ranked Feeds and Polarization
Do Recommendation Algorithms Polarize? What the Evidence Shows
Engagement-ranked feeds reward content that provokes reactions, and platforms' own researchers have raised alarms. A 2016 internal Facebook presentation, reported by the Wall Street Journal in 2020, found that 64% of joins to extremist groups it studied in Germany came through the platform's recommendation tools. But when researchers have been able to run experiments at scale, the effects of feed algorithms on political attitudes have generally been smaller than the public debate assumes.
| Study | Design | Finding |
|---|---|---|
| Guess et al. (2023), Science | Facebook and Instagram users randomly switched to chronological feeds during the 2020 US election | Less time on the platforms, but no significant change in issue or affective polarization or political knowledge over three months |
| Nyhan et al. (2023), Nature | Exposure to like-minded sources reduced by about a third | No measurable change in polarization or ideological extremity |
| Allcott et al. (2020), AER | Users paid to deactivate Facebook for four weeks | Somewhat less polarization on some measures, less news knowledge, and higher self-reported well-being |
| Hosseinmardi et al. (2021); Chen et al. (2023) | Large-scale YouTube viewing data | Consumption of extreme content concentrated among a small group of users who sought it, often via subscriptions and external links rather than recommendations |
| Eady et al. (2023), Nature Communications | Exposure to Russian influence accounts on Twitter in 2016, linked to surveys | Exposure was heavily concentrated, and was not associated with measurable changes in attitudes, polarization, or voting |
Lessons
- Optimizing a proxy (engagement) predictably shapes what gets produced and shown — that part is well established
- Effects on attitudes appear concentrated in small, already-motivated subgroups; average effects are often near zero
- Exposure is not effect: reach figures say little about persuasion
- Settle questions with experiments and data access, not with anecdotes or internal slides alone
Attention Markets
In the market-making terms of Unit 1, feeds are attention markets. Producers compete for distribution, the platform sets the rules that decide who wins, and advertising monetizes the result. The platform's ranking objective works like a price signal to producers. Reward engagement and you get more content engineered for engagement, regardless of any single user's preferences. That is why ranking objectives are policy choices even when they are framed as technical ones.
Persuasion, Targeting, and Generated Content
The persuasive power of targeted political messaging is often overstated. Across 59 randomized experiments, the effects of political advertising on candidate preferences were small regardless of context, message, sender, or recipient (Coppock, Hill & Vavreck 2020). In a pre-registered experiment, GPT-4 messages were persuasive, but microtargeted versions were no more persuasive than a single well-crafted generic message (Hackenburg & Margetts 2024).
For generative AI and misinformation, the binding constraint is usually demand and attention, not supply. There was already far more false content than anyone could read before generative models (Simon, Altay & Mercier 2023). The risks that do grow with cheap generation are specific. They include synthetic sexual imagery of real people, harassment at scale, impersonation of trusted voices, and flooding local information environments where little reliable reporting exists. The EU AI Act addresses part of this with transparency duties for AI-generated and manipulated content, applying from August 2026.
Algorithmic Public Administration
The best-documented political harms of automated decision-making come from states applying it to their own citizens. Between 2016 and 2019, Australia's 'Robodebt' scheme averaged annual tax-office income data across fortnights to infer that welfare recipients had been overpaid. It raised debts automatically and left recipients to disprove them. The 2023 Royal Commission called it a 'crude and cruel mechanism, neither fair nor legal' that 'made many people feel like criminals'. In the Netherlands, a childcare-benefits fraud system that treated nationality as a risk indicator contributed to wrongful accusations against thousands of families. The government resigned over the scandal in 2021.
Key Takeaways
- Moderation at scale requires automation; the governance task is how errors are distributed across communities and how they are corrected.
- Ranking objectives shape what is produced. Experimental evidence shows modest average effects on attitudes, with harms concentrated in subgroups.
- Exposure is not persuasion: targeting and generated content are often less potent than feared. The specific new risks are harassment, impersonation, and synthetic imagery.
- The clearest documented harms come from automated state decisions about individuals. Robodebt and the Dutch benefits scandal turned on legality, burden of proof, and recourse.
- Demand data access and experiments. Many claims in this domain — in both directions — are made without them.
Evidence & Further Reading
- Guess, A. M., Malhotra, N., Pan, J., et al. (2023). How do social media feed algorithms affect attitudes and behavior in an election campaign? Science.
- Nyhan, B., Settle, J., Thorson, E., et al. (2023). Like-minded sources on Facebook are prevalent but not polarizing. Nature.
- Allcott, H., Braghieri, L., Eichmeyer, S., & Gentzkow, M. (2020). The Welfare Effects of Social Media. American Economic Review, 110(3), 629–676.
- Hosseinmardi, H., Ghasemian, A., Clauset, A., et al. (2021). Examining the consumption of radical content on YouTube. Proceedings of the National Academy of Sciences.
- Chen, A. Y., Nyhan, B., Reifler, J., Robertson, R. E., & Wilson, C. (2023). Subscriptions and external links help drive resentful users to alternative and extremist YouTube channels. Science Advances.
- Eady, G., Paskhalis, T., Zilinsky, J., et al. (2023). Exposure to the Russian Internet Research Agency foreign influence campaign on Twitter in the 2016 US election and its relationship to attitudes and voting behavior. Nature Communications.
- Horwitz, J., & Seetharaman, D. (2020). Facebook Executives Shut Down Efforts to Make the Site Less Divisive. Wall Street Journal, May 26, 2020.
- Coppock, A., Hill, S. J., & Vavreck, L. (2020). The small effects of political advertising are small regardless of context, message, sender, or receiver. Science Advances.
- Hackenburg, K., & Margetts, H. (2024). Evaluating the persuasive influence of political microtargeting with large language models. Proceedings of the National Academy of Sciences, 121(24), e2403116121.
- Simon, F. M., Altay, S., & Mercier, H. (2023). Misinformation reloaded? Fears about the impact of generative AI on misinformation are overblown. Harvard Kennedy School Misinformation Review.
- Regulation (EU) 2022/2065 (Digital Services Act), Articles 34–35 and 40.
- Royal Commission into the Robodebt Scheme (2023). Report. Commonwealth of Australia.