POLSCI 129 Final Exam Prep

LAST ONE LAST ONE LAST ONE!1

Personalization Paradox

The promise was that tailored content would boost relevance and satisfaction (“understands” your taste, and saves your time). Optimization machinery also fostered intellectual isolation and made misinformation easier to spread. The result is a gradually narrowing worldview.

Content-Based Filtering

Recommends material similar to what you have already engaged with.

Collaborative Filtering

Recommends what other users like you have enjoyed.

Filter Bubble

A personalized digital environment shaped by your past engagement. It is driven by algorithms that prioritize engagement through content-based/collaborative filtering. Over time, your feed becomes narrow and limits your exposure to dissenting views.

Invisibility

The defining feature of the filter bubble; you are not actively blocked from other viewpoints, but you rarely encounter them. Most users do not notice this.

Echo Chamber

An environment in which people are exposed mainly to reinforcing opinions (opinions they already agree with). It is a social phenomenon rather than an algorithmic one, driven by social identity.

Homophily

The underlying mechanism of echo chambers; the tendency of people to seek out others like themselves. Even brief exposure to like-minded groups shifts individuals toward more extreme positions.

Algorithmic filtering and social homophily form a

Self-reinforcing cycle: users follow like-minded people (creating an echo chamber), and engagement-maximizing algorithms amplify that similarity. The consequence is increased polarization.

Platform Migration (after moderation)

When content moderation increased, many users migrated to alternatives like Telegram, Truth Social, and Bluesky (all of which lack fact-checking/exposure to opposing views; users form like-minded communities by choice rather than by algorithm). Echo chambers can thrive even through homophily and social identity alone. Algorithmic amplification is not the cause of the echo chamber, but an accelerant.

Is Polarization New?

Filter bubbles and echo chambers only accelerate. They speed up ideological sorting, so that what is new is not the existence of division but how quickly and how deeply users become embedded in like-minded communities.

Affective Polarization

Polarization is less about policy disagreement and more about feeling. The emphasis shifts from what the other side believes to how we feel about the other side (political sectarianism). This aversion toward the out-party has come to exceed affection for one’s own co-partisans.

Ideological Sorting Within Parties

The major parties have sorted ideologically (liberals align with Democrats and conservatives with Republicans). This alignment joins ideology with race, religion, and geography into a mega-identity. When partisanship absorbs so many identities, the other side starts to seem alien, and compromise begins to feel like a betrayal of values.

Partisan Media and Selective Exposure

As people have become more willing to consume partisan-slanted news, traditional media have become more polarized. Those inclined toward sectarian views engage in selective exposure, which reinforces their existing biases.

Polarization on Social Media

Facebook users who gave up the platform for several weeks reported lower polarization on policy issues and felt slightly happier. Emotionally charged posts are more likely to be shared within partisan networks than across them. Posts about the political out-group were shared nearly twice as often as posts about the in-group. Engagement on social media is structured around identity-based conflict: posts attacking political opponents are favored by the metrics that determine visibility.

Ideology

Measures one’s policy preferences across all issues. For elites, it is estimated from observable behavior (roll-call votes, party manifestos, endorsements, text analysis). For ordinary citizens, public opinion surveys (costly and carry problems such as social-desirability bias and nonresponse) and campaign contributions.

Birds of the Same Feather Tweet Together (Barbera)

The structure of online social networks carries information about people’s ideological positions. Under the assumption that social networks are homophilic, ideology is an underlying trait we cannot observe directly, whose value can be inferred from which political accounts each user chooses to follow. This is a costly signal that reveals information about both a user’s ideological location and that of the accounts they follow.

Rumor

A piece of information whose truthfulness is uncertain but is nonetheless widely believed and circulated. They are not a new phenomenon, but the digital era has fundamentally changed the scale/speed at which information spreads. When these claims are shared as fact, they cross over into the terrain of fake news.

Misinformation

False information spread without intent to deceive. A person might share a false headline simply because they believe it is true. There is no malicious intent, just misguided information.

Disinformation

False information shared intentionally, often for political, financial, or ideological reasons.

Fake News

The specific stories, articles, posts, and media content that serve as vehicles for spreading misinformation and disinformation. While “misinformation” and “disinformation” describe the nature and intent of false information, this describes the format through which that false information circulates (fabricated news articles designed to look legitimate, manipulated social media posts, etc).

Conspiracy Theories (False Information)

Beliefs or explanations that attribute the cause of an event or situation to a secret, often nefarious, plot by a group of powerful actors. These theories appeal to distrust in authority, provide emotionally satisfying, simplified explanations for complex events, and make believers feel as though they hold “special knowledge.”

Motivated Reasoning

A tendency to believe information and narratives that align with our existing beliefs and political identity. This tendency is driven by several specific cognitive biases that work together to shape what we are willing to accept as true, like belief alignment (accept false stories that are consistent with their beliefs and to reject true stories that contradict them), In-group/Out-group Thinking (believe positive information about one’s own group and negative information about the opposing side), and Confirmation Seeking (seek out and favor information that confirms their existing beliefs, which strengthens their initial stance).

Effective Messengers

The messenger matters more than the message (also called an Unexpected Validator); people speaking against their own apparent group or interest are the most effective correctors; repeating a lie to debunk it can work (contrary to the Backfire Effect).

Telling the Truth About Believing the Lies (Berinsky)

Do people honestly report believing a rumor? There are two mechanisms: Genuine Belief (no conscious effort to deceive the surveyor; people who actually believe false information may oppose policies they would otherwise support, and such a belief can undermine trust in political actors and institutions), and Cheap Talk (intentionally misrepresents their personal belief; the opinion expressed differs from the opinion actually formed).

(Berinsky) Experiments on two rumors (that Obama is Muslim and that the government knowingly allowed 9/11) showed that

Belief barely changes when incentives (spending less time on the survey, etc) change, suggesting genuine belief, not just cheap talk.

Because of selective exposure and confirmation bias

Genuine belief may not change votes (it mostly confirms existing views), but rumors still harm democracy by undermining trust and intensifying polarization.

Fact-Checking Red Flags

Sensational wording, emotional appeals, clickbait structure, though some fakes are subtle.

Cognitive Biases Behind Motivated Reasoning

Belief Alignment: People are inclined to accept false stories that are consistent with their beliefs and to reject true stories that contradict them.

In-group/Out-group Thinking: A willingness to easily believe positive information about one’s own group and negative information about the opposing side, while rejecting negative information about one’s own group.

Confirmation Seeking: People tend to seek out and favor information that confirms their existing beliefs, which strengthens their initial stance.

Social Network

The complex web of relationships that connect individuals. Unlike popular platforms such as Facebook or Instagram, social networks in the academic sense refer to the actual human connections that exist between people in the real world (relationships from a person’s geography, family, work, and activities; form the foundation of our social lives and influence aspects of our behavior, like our political choices).

Political Networks

A subset of our broader social networks, consisting of members with whom an individual discusses politics, elections, or government matters. Typically dominated by primary group members, people with personal connections to us (family members, close friends, and trusted colleagues). These are the people we feel comfortable discussing potentially sensitive political topics with.

Nodes

Entities (individuals, organizations, countries) that can form relationships with other entities. In a social network, they typically represents people.

Edges/Links

Connections between nodes: friendships, family relationships, professional collaborations, or who knows whom.

Types of Network Structures

Density (proportion of actual connections to possible connections; dense networks tend to spread information quickly), centralization (whether connections are concentrated around a few key nodes (centralized) or spread more evenly across the network (decentralized)), and clusters/bridges (clusters are subgroups where nodes are tightly interconnected; bridges are nodes or ties that connect otherwise separate clusters).

Links can be ___, where the relationship does not go both ways (one person follows another on Twitter, but not vice versa), or ___, where the direction does not matter (two people are friends, which is mutual).

Directed; Undirected

Some nodes hold more strategic or influential positions than others. We measure this using centrality metrics:

Degree Centrality: The number of links held by each node. It shows how many connections a node has.

Betweenness Centrality: The number of times a node lies on the shortest path between other nodes. It identifies “bridge” nodes that connect different parts of the network.

Closeness Centrality: Scores each node by its “closeness” to all other nodes, calculated from the shortest paths between all nodes.

The measures can identify different important nodes in the same network: the node with the most direct ties (degree) is not always the one that bridges separate clusters (betweenness) or the one that can reach everyone else most efficiently (closeness).

Renaissance Florence

In 15th-century Florence, political power was often built not through battles but through strategic marriages. Powerful families like the Medici used marriage ties to place themselves at the center of Florence’s political life. The Medici’s marriage ties, not their wealth, made them central; the richer-but-less-connected Strozzi lost power.

Political networks tend to be

Low-density (in very few cases are an individual’s political discussion partners socially connected) and asymmetric (you may identify someone as a political discussant, but that person may not reciprocate).

Political networks influence individuals through two primary mechanisms:

Information Sharing: Political networks expose us to information, arguments, and interpretations that might not otherwise reach us.

Social Pressure and Conformity: We feel pressure to align our views and behaviors with those of our social connections. This pressure is stronger when the behavior is publicly visible and when our connections are monitoring our behavior. (Classic get-out-the-vote experiments that apply social pressure, e.g., telling neighbors who did and did not vote, illustrate this mechanism).

One of the hardest problems in network analysis is determining causality. When we observe similar political behaviors or opinions within a network, two explanations are possible:

Network Influence: People in the network influenced each other through conversation and social pressure.

Homophily (“birds of a feather flock together”): People selected network members who were already similar to themselves, whether in political preferences or in socioeconomic and demographic characteristics correlated with politics.

Political Networks in the Digital Age

Voter mobilization (Facebook “I Voted” networks), hashtag activism (#MeToo and #BlackLivesMatter), polarization, influence networks (political influencers shape discourse), misinformation spread, and transnational activism (protesters coordinate across borders using social media).

How Political Networks Influence Individuals

Information sharing (political networks expose us to information, arguments, and interpretations that might not otherwise reach us) and social pressure/conformity (we feel pressure to align our views and behaviors with those of our social connections).

Researchers use several strategies to avoid overstating the influence of networks, such as

Randomized field experiments (introducing a mobilization stimulus/tracking it through a social network lets researchers observe the spread of a new behavior), panel data (tracking individuals’ political preferences over time can show convergence with network members’ views), and statistical controls (controlling for the social characteristics that generate homophily, like by comparing similar pairs of people with and without a social connection).

Locating and Measuring Networks

Geography as proxy (objective data: all members within a particular geography form a network), behavioral connections (objective data: identify network members through shared actions), and self-reported networks (subjective data: ask individuals to identify the members of their political network).

Coordination Problems (Collective Action Problem)

Getting large numbers of people to act together requires overcoming logistical hurdles: agreeing on the timing, location, and method of action, and disseminating the information people need to act in concert, such as transportation, meeting points, police presence, medical services, and legal support. The classical solutions are leadership and geographic focal points (for example, protesting in a well-known public square), which give people a natural place and time to converge.

Cooperation Problems (Collective Action Problem)

Even when coordination is solved, movements must motivate individuals to participate when the action carries personal costs or risks. The difficulty is that the benefits of success (say, a policy change) are experienced collectively, by participants and non-participants alike, while the costs (risk of arrest, injury, lost wages, time) are borne individually. This is the free-rider problem: a rational individual is tempted to let others bear the costs while still enjoying the benefits, which can leave a movement with too few participants to succeed.

A central classical solution to the free-rider problem is the selective incentive, or

A reward that can be experienced individually, beyond the collective goal, such as social approval, a sense of belonging, reputation, or other psychological rewards. Because selective incentives accrue only to those who participate, they change the personal cost-benefit calculation and make cooperation rational. Movements have long created them through strong social bonds, group norms of approval and disapproval, and the psychological satisfaction of contributing to a valued cause or expressing one’s identity.

Internet Surveillance

The use of technology by government agencies to monitor the online activity of individuals or groups. Its defining feature is monitoring: governments watch, collect, and analyze communications and behavior. People may not know they are being watched, and their ability to access information is largely unchanged in the immediate term.

Internet Censorship

The use of technology by government agencies to suppress access to certain websites or apps, or the publication of certain materials. Its defining feature is restriction: governments actively prevent access or expression. Operates visibly, as users encounter blocked sites, removed content, or restricted capabilities. It directly inhibits freedom of expression and is generally seen as a restriction on internet freedom.

Censored: Distraction and Diversion Inside China’s Great Firewall (Margot Roberts)

Bypassing censorship takes time, effort, and technical knowledge. Even when circumvention is technically possible, these costs create “friction” in the information market. Think of it as adding a toll to a previously free highway: some people pay the toll, but others take alternative routes or travel less. The mere presence of added cost changes overall traffic patterns, even though the road remains technically open.

The Three Mechanisms of Censorship (the “3 Fs”)

Fear: People are deterred from reading or writing certain material out of fear of punishment, and so self-censor. This is highly visible but risky for governments, because harsh tactics can backfire by generating sympathy for victims, increasing public interest in the banned content, and depriving authorities of useful information about public sentiment.

Friction: making information more difficult or costly to access or share, without fully blocking it. It relies on human tendencies toward convenience and impatience, separating the casual “masses” from determined activists. China’s “Great Firewall” is primarily a friction mechanism.

Flooding: exposing people to a noisy, distracting information environment in which it is hard to judge the importance or relevance of any piece of information. Rather than removing content, flooding dilutes it by adding large volumes of alternative material, exploiting limited cognitive capacity.

China Study (King, Pan & Roberts)

1,382 sites in the first half of 2011; censors act fast (within about 24 hours). Posts with collective-action potential are censored the most (about 80%), while policy and news posts are censored rarely (about 10%), whether they support or criticize the state.

Collective Action Theory

Authoritarian governments are primarily concerned not with preventing criticism but with preventing the organization and coordination that criticism might facilitate. Individual complaints can actually be useful to a regime, providing information about local problems and an outlet for frustration. What threatens stability is not criticism itself, but criticism becoming organized into collective challenges.

Internet Censorship in the United States

Legal takedowns (DMCA), National Security Letters with gag orders, CIPA institutional filtering, and private platform moderation.

Platform Moderation

Meta began eliminating professional fact-checkers and replacing them with user-generated “community notes” similar to X’s system. Zuckerberg argued that fact-checkers had “destroyed more trust than they’ve created” and that inclusion efforts had gone too far in shutting down differing opinions, a fundamental shift in how a major platform approaches moderation, reflecting political pressures and evolving expectations.

Santa Clara Principles

Establish basic expectations: respect for human rights; clear rules users can understand without legal expertise; cultural competence that accounts for linguistic and contextual differences; transparency about government influence on content decisions; and due process, including clear notification when content is restricted.

For major platforms, operational principles set more demanding expectations: regular transparency reports with detailed statistics; meaningful appeals processes reviewed by humans who can consider context; and transparency about how automated tools are used.

TikTok v. Garland (2025)

The 2024 divest-or-ban law was unanimously upheld under intermediate scrutiny, framed as regulating ownership, not speech, establishing digital sovereignty. Aftermath: brief suspension, enforcement delays, and a sale closing January 23, 2026 (Oracle-led ∼80%, ByteDance under 20%).

Digital Resistance

VPNs, Tor, and Signal; code words, euphemisms, emoji, and meme warfare; cross-border support networks. But the same tools that enable legitimate resistance can facilitate harm, the “dual-use” problem: VPNs that protect journalists also hide criminals; encryption that protects dissidents can enable extremist coordination; creative evasion that defeats political censorship can also spread misinformation or harassment.

Why does AI Matter for Politics?

Campaigns use it to target, personalize, and automate political messaging. Governments use it to deliver services, track engagement, and inform policy decisions. And voters encounter it every time they search, scroll, or get served a political ad, usually without realizing it.

From Learning to Generating

Political AI was overwhelmingly about learning and prediction. Machine-learning systems studied patterns, what you click, who you follow, how long you watch, in order to predict what you would want to see next. Then AI learned to create. Generative AI produces new content, text, images, audio, and video, that can be effectively indistinguishable from the real thing. The consequences for campaigns are immediate: a campaign can now clone a candidate’s voice, generate a polished attack ad without a film crew, and simulate an endorsement that never happened.

Cheap

The capacity to fabricate convincing political media is no longer limited to states or studios; it is available to any campaign, troll, or individual.

Thin

The line between real and fake has become thin, which corrodes a precondition of democratic argument: a shared ability to tell what actually happened.

The New Propaganda Machine

AI has industrialized political manipulation. A single individual can now flood social media with thousands of coordinated, human-sounding accounts. Those bots can manufacture a fake consensus: a fringe idea wrapped in 10,000 fake engagements looks mainstream, and “looks mainstream” is often enough to become mainstream. The deepest danger is not any single false post. It is that when fake and real become indistinguishable, trust in everything erodes, including legitimate information. The practical upshot is unsettling: you can no longer assume that the political conversation you are watching online reflects what real people actually think.

Trump and AI

Donald Trump is the first U.S. president to use AI-generated imagery as a routine tool of political communication, mostly without disclaimers. A partial timeline from the deck February 2025, “Trump Gaza,” depicting Gaza transformed into a luxury resort bearing his name; May 2025, Trump dressed as the Pope, posted days after Pope Francis died; October 2025, a fighter jet dumping waste on “No Kings” protesters. Earlier and later examples extend the pattern: in 2024, AI images falsely showed Taylor Swift endorsing Trump (Swift later said this motivated her real endorsement of Harris); in September 2025, hours after meeting Democrats at the White House, he posted an AI video mocking them; and in April 2026, an AI image of Trump as Jesus appeared hours after a post criticizing Pope Leo XIV.

Deepfakes of Candidates

The clearest illustration is the Talarico case (March 2026): the National Republican Senatorial Committee posted an 85-second AI-generated video of Texas Democratic Senate candidate James Talarico appearing to read and endorse his own controversial tweets. The fabricated “Talarico” was hyper-realistic; experts said most viewers would not detect it. A small “AI GENERATED” label appeared in the corner, which critics called easy to miss. Crucially, Texas law bans political deepfakes only within 30 days of an election, and the ad fell outside that window, and no federal law regulates AI in political ads at all.

Global Cases of AI

Germany (AfD videos), Ecuador (deepfake anchors), Canada (Carney deepfakes), Argentina, South Korea (AI avatars), France (Macron), and Ghana/ Taiwan / India (bots and foreign influence).

Manufactured Consensus (Astroturfing)

Fake grassroots support simulated at scale, often via bots, to make a view look popular.

Political Bot

An automated account that mimics a human to amplify, persuade, or drown out opposition.

Microtargeting Tailoring

Messages to narrow audiences using behavioral, social, and psychological data.

Liar’s Dividend

The advantage a liar gains when the existence of deepfakes lets them dismiss genuine evidence as fake.

Disclosure

The question of whether and how machine-authored political content must be labeled.

Regulatory Vacuum

The patchwork of inconsistent state rules and absent federal law governing AI in political advertising.