Social Media and Elections: How Algorithms Shape Votes | State of Surveillance

TL;DR: Social media platforms don't just host political content — they actively shape what you see through recommendation algorithms optimized for engagement, not accuracy. Cambridge Analytica showed how psychological profiling could target voters with personalized propaganda. Bot networks flood platforms with fake activity to manipulate trending topics. Deepfakes are becoming indistinguishable from real footage. Research suggests about 25% of social media interactions in some domains are driven by bots. A 2025 study found that just one week of algorithmic content curation shifted users' political attitudes by an amount normally seen over three years. The platforms know their systems can be weaponized. They've chosen scale over safety.

How Social Media Algorithms Work

Every major social media platform uses recommendation algorithms to decide what content you see. These systems analyze your behavior — likes, shares, comments, time spent viewing — and predict what will keep you engaged.

The goal isn't to inform you. It's to keep you scrolling.

The Engagement Optimization Problem

Recommendation algorithms learn what captures attention. Research consistently shows that emotionally provocative content — especially content that triggers fear, anger, or outrage — generates more engagement than neutral information.

This creates a feedback loop:

Outrage-inducing content gets more engagement
Algorithms surface content with high engagement
More users see outrage-inducing content
Users who engage get shown more extreme content
The algorithm learns that extreme content "works"

Political content is particularly susceptible. Partisan attacks, conspiracy theories, and inflammatory claims generate more clicks than nuanced policy discussions. Algorithms don't distinguish between "engaging because it's informative" and "engaging because it's inflammatory."

YouTube's 70% Problem

YouTube's recommendation algorithm is responsible for roughly 70% of what users watch on the platform. Users don't choose most of what they see — the algorithm chooses for them.

Researchers using "sock puppet" accounts to study recommendations found that YouTube amplifies extreme and fringe content through algorithmic recommendations. An account that watches one conspiracy video gets recommended more. An account that watches political content gets pushed toward increasingly partisan sources.

The same pattern appears across platforms. TikTok's algorithm, in experiments during the 2024 U.S. presidential election, showed measurable political bias — Republican-seeded accounts received approximately 11.8% more party-aligned recommendations compared to Democratic-seeded accounts.

Filter Bubbles and Echo Chambers

In 2011, internet activist Eli Pariser coined the term "filter bubble" to describe how algorithmic personalization isolates users in information silos. If you interact with conservative content, you see more conservative content. Progressive content disappears from your feed. Eventually, you're only seeing one perspective — and you don't even know what you're missing.

Echo chambers are related but distinct: they form when users choose to engage primarily with like-minded people and sources. The algorithm then reinforces this choice by showing more of what you already agree with.

What Research Shows

The evidence on filter bubbles is more nuanced than the popular narrative suggests. A comprehensive literature review from the Reuters Institute found that echo chambers are "much less widespread than is commonly assumed" and that digital media may increase "perceived rather than actual polarization."

However, this doesn't mean algorithms are harmless. Research confirms that:

Highly polarized or radicalized groups do exist in tight echo chambers
Algorithms can accelerate exposure to extreme content for users who start engaging with it
Even when users encounter opposing views, they often react with hostility rather than consideration
The overall effect is to sort users into increasingly homogeneous communities

A 2025 study published in Science installed browser extensions on consenting participants to automatically rerank their social media posts. After just one week, users' feelings toward the opposing political party shifted by about two points on a standard scale — an effect normally observed over three years.

That's the power of algorithmic curation: measurable attitude change in days.

Cambridge Analytica: The Microtargeting Playbook

In 2018, whistleblowers revealed that Cambridge Analytica — a political consulting firm — had harvested data from up to 87 million Facebook profiles without meaningful consent. The company used this data to build psychological profiles of American voters, then delivered targeted political messages designed to exploit individual psychological vulnerabilities.

How It Worked

Data collection: A third-party Facebook app collected data not just from users who installed it, but from all their Facebook friends — including status updates, likes, and private messages
Psychological profiling: Using the "Big Five" personality model (openness, conscientiousness, extraversion, agreeableness, neuroticism), the company categorized users into personality types
Tailored messaging: Different ads were shown to different personality types, designed to trigger specific emotional responses
Emotional exploitation: Messages were crafted to exploit fear, anger, or hope depending on what would be most effective for each individual

By 2017, Cambridge Analytica claimed it had psychological profiles of 220 million U.S. citizens based on 5,000 separate data sets.

The "Do So" Campaign

One documented example reveals the company's willingness to suppress votes. In 2010, Cambridge Analytica designed a campaign called "Do So" aimed at increasing abstention among young people of African descent. The campaign was presented as if it arose spontaneously on social media as a resistance movement encouraging young people not to vote as a form of protest.

It wasn't organic. It was manufactured voter suppression.

Did It Work?

The effectiveness of Cambridge Analytica's techniques remains debated. Some research suggests the company exaggerated its capabilities — Trump campaign aides disputed the company's role, describing it as "modest." The New York Times reported in 2017 that Cambridge Analytica executives conceded the company "never used psychographics in the Trump campaign."

But the technique itself is real. Whether Cambridge Analytica succeeded, the playbook exists. Whistleblowers from the company claim it engaged in psychological operations across 68 countries before folding in 2018.

A British Parliamentary investigation concluded that relentless targeting that "plays to the fears and the prejudices of people, in order to alter their voting plans" is "more invasive than obviously false information" and contributes to a "democratic crisis."

Bot Networks: Manufactured Consensus

Research estimates that approximately 25% of social media interactions in some domains are driven by bots. Globally, nearly half of all web traffic is non-human. These aren't just spam accounts — many are sophisticated influence operations.

How Bot Networks Operate

Modern bot operations use coordinated networks of fake accounts to:

Amplify content: Hundreds of accounts like, share, and comment on the same post to trigger algorithmic promotion
Create trending topics: Coordinated posting makes a hashtag appear organically popular
Suppress opposition: Mass reporting of legitimate accounts can trigger automated suspensions
Spread disinformation: Multiple "independent" accounts sharing the same false story creates an appearance of credibility
Manufacture consensus: If everyone in your feed seems to agree on something, you're more likely to agree too

Researchers found accounts that flood networks with tens or hundreds of thousands of posts in a single day. The same campaign can post a message with one account and then have other accounts "like" and "unlike" it hundreds of times in a short time span, then delete the messages to evade detection.

Foreign Operations

The 2024 U.S. election faced sophisticated foreign influence campaigns from state-sponsored actors including Russia, China, Iran, and Israel. The Department of State announced sanctions against both the Cognitive Design Production Center (a subordinate of Iran's Islamic Revolutionary Guard Corps) and the Russia-based Center for Geopolitical Expertise for interference in the 2024 election.

Meta found and disabled thousands of fake Facebook accounts linked to China that were used to spread politically partisan content. These accounts reshared posts from both liberal and conservative sources — the goal wasn't to support one side but to "exaggerate partisan divisions and further inflame polarization."

In a notable 2025 case, an AI deepfake of Canadian Prime Minister Mark Carney reached more than one million views on social media before the election.

Deepfakes and AI-Generated Disinformation

Benchmark comparisons confirm that AI-fueled disinformation is accelerating. Researchers documented a fivefold rise in deepfakes and tens of millions of fake identities since 2020.

In the 2024 Indian regional elections, deepfake videos were deployed to subvert language barriers and spread targeted propaganda. Partisan operatives created videos of political leaders mouthing words in languages they never spoke. Over 15 million people were reached via approximately 5,800 WhatsApp groups that circulated these deepfakes.

The 2024 U.S. Election

Research evaluating the impact of AI-generated disinformation on U.S. election integrity found no conclusive evidence that such content manipulated the results. However, deceptive AI-generated content still influenced election discourse, amplified harmful narratives, and entrenched political polarization.

The capacity exists. The tools are cheap and accessible. Low-quality deepfakes can be created in minutes. High-quality ones require more effort but are increasingly within reach of non-experts.

What Makes Deepfakes Dangerous

Speed: A false video can go viral before fact-checkers respond
Emotional impact: Video is more persuasive than text
Doubt as a weapon: Even real footage can be dismissed as fake
Scale: AI enables personalized disinformation at mass scale

As of early 2025, an estimated 5.64 billion individuals — approximately 68.7% of the world's population — were active internet users. This level of global digital saturation offers a fertile environment for disinformation to propagate rapidly.

How Platforms Have Responded

Social media companies have introduced various policies to address election manipulation. Their effectiveness is questionable.

Meta (Facebook, Instagram)

Requires political advertisers to disclose AI-generated content (effective January 2024)
Blocks new political ads during the final week before U.S. elections
Labels state-controlled media outlets
Claims to have removed over 200 malicious influence operations

However, a ProPublica investigation found that of more than 110,000 ads on deceptive "Patriot Democracy" pages, Meta stopped just over 7,000 — roughly 6% — from running for violating standards. These ads were shown nearly 60 million times before Meta took action. Meta also consistently failed to detect and remove copies of ads it had previously banned.

Advocacy groups note that Meta, YouTube, and X rolled back 17 policies intended to protect against hate speech and misinformation in the lead-up to 2024.

TikTok

TikTok has stated its commitment to balancing freedom of expression and community safety. But its moderation policies and enforcement processes remain opaque.

Analysis of 51,680 political videos from the 2024 U.S. presidential cycle revealed that toxic and partisan content consistently attracts more user engagement despite moderation efforts. Posts about immigration and election fraud drew particularly high levels of toxicity.

The Fundamental Problem

Platforms face a structural conflict: their business model depends on engagement, and inflammatory content drives engagement. Every policy that reduces the spread of disinformation also reduces time spent on the platform.

They've chosen their revenue model. Everything else is damage control.

Regulatory Responses

Governments are beginning to respond, though approaches vary dramatically.

European Union

The EU adopted Regulation 2024/900 on the transparency and targeting of political advertising. This regulation:

Harmonizes rules across EU member states on how political ads must be labeled, tracked, and disclosed
Sets limits on using personal data for targeting
Establishes oversight mechanisms

The regulation takes effect in October 2025.

United Kingdom

As of early 2025, the UK relies on general data protection law, consumer protection law, and voluntary industry measures. Critics worry this decentralized approach may lag behind fast-moving threats like synthetic media manipulation.

United States

Representatives Bill Foster and Tom Kean, Jr. sent a letter to social media platforms demanding action against foreign bot farms spreading disinformation and encouraging political violence. But comprehensive federal legislation remains absent.

The FEC has proposed requiring disclosure of AI-generated content in political ads, but enforcement mechanisms remain weak.

What This Means for Voters

If you use social media, your political environment is being shaped by:

Algorithms that prioritize engagement over accuracy
Microtargeting that delivers different messages to different people based on psychological profiles
Bot networks that manufacture the appearance of consensus
Deepfakes that can put words in anyone's mouth
Foreign operations designed to inflame division

The information environment you experience is not the same as what your neighbor experiences. You may be seeing entirely different "realities" constructed by algorithms and influence operations — and neither of you knows what you're missing.

The Perception Gap

Research suggests that digital media may increase "perceived rather than actual polarization." Americans believe the other side is more extreme than it actually is. Social media amplifies the loudest, most inflammatory voices and makes them seem representative.

This perception gap is itself politically useful. If you believe the other side is full of extremists, you're more likely to support extreme measures against them. The manufactured perception becomes a self-fulfilling prophecy.

What You Can Do

Complete protection from algorithmic manipulation isn't possible for anyone who uses social media. But you can reduce your exposure:

Diversify Your Sources

Deliberately seek out news sources you disagree with
Follow journalists, not just commentators
Read primary sources when possible (court documents, official statements, research papers)
Use RSS readers to bypass algorithmic curation

Recognize Manipulation Tactics

Content designed to trigger outrage is often designed to be shared, not to inform
Claims that "everyone" agrees on something should raise suspicion
Verify before sharing — reverse image search, check the source, wait for confirmation
Ask: "Who benefits if I believe this?"

Limit Algorithmic Exposure

Use chronological feeds when available (they're usually buried in settings)
Clear watch history and recommendation data periodically
Use incognito/private browsing for political content to prevent profile building
Consider reducing social media use during election seasons

Verify Deepfakes

Look for unnatural blinking, lighting inconsistencies, or audio-video sync issues
Check if the content appears on the original source's official channels
Wait before sharing — breaking news is often wrong news
Remember that even real footage can now be dismissed as fake

The Bottom Line

Social media platforms have become the primary information environment for billions of people. The algorithms that control what those people see are optimized for engagement, not truth. The same systems that keep users scrolling are the systems that hostile actors exploit to manipulate elections.

Cambridge Analytica proved that psychological profiling and microtargeting work — even if that company exaggerated its own role. Bot networks can manufacture the appearance of grassroots movements. Deepfakes can put words in anyone's mouth. Foreign governments are actively using these tools to undermine democratic processes.

The platforms know this. They've known it for years. They've chosen not to fundamentally change their business models because engagement-driven algorithms are profitable.

Until that calculus changes — through regulation, competition, or user revolt — social media will remain a weapon as much as a tool. The question isn't whether these systems are being used to manipulate elections. The question is how effectively, and by whom.