Computational Social Science

Social science used to mean surveys, interviews, and a few hundred people in a lab. Computational social science uses the traces billions of us leave: posts, clicks, buys, GPS, search. ML instead of a simple regression. Graphs instead of a focus group. Platform logs instead of a clipboard. Millions of rows, not hundreds. The catch is the same as the gift: APIs close when a company gets bored, the feed you see is already filtered, and the sample is not "everyone." You can watch more behavior than any prior method. You can also lose the firehose overnight.

Tap to enlarge

The data source problem

For two decades, Twitter (now X) was the primary data source for computational social science. Its API provided researchers with near-real-time access to public posts for studies of political polarization, misinformation spread, disaster response, public health sentiment, and social movement dynamics. An estimated 25,000+ peer-reviewed papers were built on Twitter data, making it the single most studied platform in the history of social science.

In 2023, following Elon Musk's acquisition, Twitter imposed API pricing that effectively ended free academic access. The academic tier, which had provided researchers with up to 10 million tweets per month at no cost, was eliminated. The replacement pricing, approximately $42,000 per month for comparable access, was prohibitively expensive for most university research budgets. Reddit followed a similar path in mid-2023, shutting down the widely used Pushshift archive and introducing API pricing at $0.24 per 1,000 calls.

The consequences for the field are structural. Longitudinal studies that tracked political behavior, public health attitudes, or information diffusion over years lost their data stream overnight. Replication of prior findings became difficult or impossible. The platforms that had been a public commons for social observation became private resources available primarily to well-funded corporate and government entities.

The disruption forced the field to see a dependency it had treated as a public good. CSS had built its methods on the assumption that platform data access was stable. It was a private resource that a single company could revoke at any time, for any reason.

The methodological evolution

The loss of easy API access forced rapid methodological innovation. The methods that replaced API scraping often measure things the API never showed. Each has its own limits.

Browser extension studies. Researchers at NYU, Princeton, and other institutions now recruit volunteer participants who install custom browser extensions. These extensions log the content participants see in their social media feeds, what they post, and what the algorithm surfaces. This provides ground truth about algorithmic exposure that API data never offered: the old approach captured what users said, but not what they were shown. The limitation is scale: browser extension studies typically involve thousands of participants rather than the millions available through APIs, and participants are self-selected.

Data donation. GDPR Article 15 and equivalent regulations give individuals the right to export their complete interaction history from any platform. Academic protocols now guide participants through downloading their data packages, which include likes, saves, dwell time, scroll behavior, search history, and algorithmic recommendations, and sharing them with researchers under informed consent. This produces richer individual-level data than API scraping ever provided, though at far smaller scale and with significant effort required from each participant.

Digital twins. Researchers are using Large Language Models to create synthetic replicas of user personas: "digital twins" trained on real user-generated content that simulate how individuals might respond to different stimuli. A 2024 Stanford study demonstrated that GPT-4-based agents could replicate individual survey responses with 85% accuracy when calibrated on a person's social media history. These synthetic agents can model scenarios (policy changes, platform design modifications, information campaigns) without exposing real users to experimental manipulation. The limitation is validation: how do you confirm that synthetic behavior accurately represents real behavior without the real behavioral data you no longer have access to?

Algorithmic auditing. Rather than studying user behavior, a growing body of research studies algorithm behavior. Researchers create controlled "sock puppet" accounts with different disclosed demographics (age, gender, political orientation, location) and measure divergence in algorithmic treatment: what content is recommended, what is suppressed, what advertisements are shown. This approach treats the algorithm itself as the subject of study, not users, and has produced some of the field's most important findings about systematic amplification and suppression patterns.

Platform partnerships. The 2023 Meta/Science collaboration, which produced four papers on algorithmic polarization, is the partnership model: researchers work with platform companies under NDA to access internal data at scale. The advantage is data APIs never expose, including dwell time and scroll velocity. The disadvantage is corporate control. Meta retained the right to review findings before publication, and critics have noted that the research design, which focused on short-term experimental manipulation of individual features, was structured in a way that was unlikely to find large effects.

The polarization question

The largest-scale CSS studies have focused on political polarization, with results that challenge the popular narrative that "social media causes polarization."

The 2023 Meta/Science collaboration experimentally modified the Facebook and Instagram feeds of 23,000 users during the 2020 US election. The research produced four papers, each testing a different intervention:

Chronological feed experiment. One group saw chronologically ordered content instead of algorithmically ranked content. The algorithmic feed increased exposure to ideologically aligned content and decreased exposure to cross-cutting perspectives. However, switching to chronological feeds did not measurably reduce political polarization in attitudes. Users exposed to chronological feeds spent less time on the platform, clicked less, and reacted to fewer posts.

Reshare removal experiment. Removing reshared content from Facebook feeds significantly reduced the volume of political news and misinformation that users encountered. The reshare mechanism, which allows content to propagate through networks at low cognitive cost (one tap), is a primary vector for both political content amplification and misinformation spread. This finding is the most actionable of the four. Reshares are a design choice, and removing them changed the information mix.

Like-minded source reduction. Reducing the proportion of content from like-minded sources decreased exposure to news from untrustworthy sources. Political news consumption is highly ideologically segregated, with users seeking out and engaging more with content that aligns with their pre-existing views. The echo chamber effect is real, and the algorithm maintains it.

These findings are frequently misquoted. A common reading is that algorithms do not cause polarization. The papers show something narrower: short-term experimental manipulation of a single platform feature does not rapidly change political attitudes that have been forming for years. The long-term, cumulative effect of years of algorithmic curation remains open. These studies, by design, could not measure it.

The representation problem

A limitation of CSS is population representativeness. No social media platform constitutes a representative sample of any national population.

Twitter/X's user base skews younger, more male, more urban, more politically engaged, and more socioeconomically advantaged than the general population. Facebook's user base skews older. TikTok's skews dramatically younger (Gen Z and Gen Alpha). Reddit's skews male and tech-literate. LinkedIn's skews professional and affluent. Each platform selects for a different subset of the population, and the design of each platform shapes the behavior it observes.

"Twitter sentiment" is the sentiment of people who express opinions on Twitter, filtered through an algorithm that amplifies emotionally charged content. Research that draws conclusions about "public opinion" from any single platform's data is drawing conclusions about a non-representative subset of the public, expressed in a format the platform's design incentivizes, and filtered through an algorithm optimized for engagement.

The observer effect operates at population scale. Social media platforms are observation environments that shape the behavior they record. Their design incentivizes specific behaviors: brevity (character limits on Twitter), emotional intensity (engagement algorithms reward outrage), public performance (visibility metrics reward display), and conformity (social proof through likes and shares discourages minority viewpoints). Behavior observed on these platforms is behavior shaped by these platforms. The instrument of measurement alters what is being measured.

The ethics infrastructure

CSS raises ethical questions that traditional social science frameworks were not built to handle.

The Facebook emotional contagion study (2014) remains the defining ethical controversy. The study experimentally manipulated the emotional content in 689,000 users' news feeds without informed consent, demonstrating that emotional states are contagious through social networks: users exposed to more negative content posted more negatively themselves. The backlash targeted the method: manipulating the emotional experiences of nearly 700,000 people at scale, without their knowledge or consent, under the legal cover of Terms of Service rather than Institutional Review Board (IRB) approval.

The ethical standards that have evolved since then remain inconsistent:

Institutional Review Boards (IRBs) at universities now require explicit informed consent for social media experiments involving any form of manipulation or observation of identifiable individuals. A graduate student studying 500 tweets needs IRB approval. But IRB jurisdiction does not extend to corporate research. Platform companies can conduct internal experiments on their own users, at any scale, under their Terms of Service. Facebook alone runs thousands of A/B tests per day on its user base, each one a social experiment conducted without informed consent.

Data protection regulations (GDPR in Europe, CCPA in California) provide individuals with legal rights over their data, including the right to access, export, and delete it. These regulations have created the legal foundation for data donation research but leave the power asymmetry in place: platforms hold fine-grained behavioral data, and researchers must negotiate access on the platform's terms.

The "do no harm" principle extends to publication effects. In polarized environments, even publishing research about group behavior can be weaponized. Findings about the online behavior of a political, ethnic, or religious group can be cited, often out of context, to justify discrimination, surveillance, or platform-level suppression. CSS researchers increasingly face decisions about whether publishing accurate findings is in the public interest or provides ammunition to bad actors.

The emerging infrastructure

The field is building methods that do not depend on the goodwill of platform corporations.

Decentralized platforms. Bluesky (built on the AT Protocol) and the Mastodon / ActivityPub federation provide open, researcher-accessible data streams by design. Bluesky's "firehose" API provides real-time access to the entire public post stream at no cost. Mastodon's federated architecture allows researchers to access instance-level data with server administrator consent. These platforms are small relative to incumbents, but they are growing, and their architecture ensures that data access cannot be unilaterally revoked.

Government-mandated access. The EU's Digital Services Act (DSA), which took full effect in 2024, requires very large online platforms (VLOPs) to provide researchers with access to data for studying systemic risks. Article 40 establishes a framework for "vetted researcher" access, potentially creating a legal right to platform data for qualified academic researchers within the EU.

Synthetic data and simulation. LLM-based social simulations, where populations of AI agents interact under controlled conditions to model social dynamics, are emerging as a complement to observational research. These "silicon societies" allow researchers to test hypotheses about collective behavior at scale, without the ethical constraints of human experimentation and without dependency on platform APIs.