Computational Social Science Methods
The Closing of the Open Web and the Politics of Data Access
A recurring spine across this corpus is the diagnosis that computational social science is living through — and methodologically responding to — the collapse of the open-API era. Freelon2024-sc periodizes this trajectory from laissez-faire access through the post-Cambridge Analytica clampdown to today’s fragmented patchwork of pay-to-play, walled-garden, and unofficial regimes. Murtfeldt2025-wu supplies the bibliometric evidence: a “Golden Age” of Twitter research that grew at ~25% annually through 2022, then stagnated and declined once free APIs disappeared. Yang2026-tq generalizes this beyond Twitter, showing that the share of social-science papers using social media data tripled between 2010 and the early 2020s before plateauing — concentrated on Twitter/X and Facebook even as users moved to YouTube and TikTok. The political stakes of this contraction are framed normatively by Heiss2026-qv and empirically by Bak-Coleman2026-mk, who finds ~80% “industrial saturation” of high-profile social-media research through undisclosed author, editor, and reviewer ties. Peters2026-mo then takes the regulatory turn, showing how data quality became a contested object in the EU DSA Article 40 consultation process, with academics and NGOs partially succeeding in writing their methodological concerns into the Delegated Regulation.
Auditing Platforms from the Outside
If platforms will not open up, researchers must audit them. Rieder2025-ju dissects YouTube’s search API as “forgetful by design,” with severe temporal decay and inconsistent results across identical queries — calling into question retrospective research designs the field has taken for granted. McNally2025-dn and its companion Bastos2025-ya (and Bastos2025-ol) reject the “black box” metaphor for Facebook’s News Feed, using 52 documented ranking updates and time-series methods to recover a ~19–24-day lagged effect on hard-news engagement. Bouchaud2026-lr performs perhaps the most ambitious external audit, reconstructing X’s user embedding space from 2.5M donated “Who to Follow” recommendations and showing that ideology is encoded as a linear direction (ρ=0.887) — turning recommender auditing into a question about internal representations rather than just outputs. Inacio-da-Silva2026-zf anticipates this volunteer-data strategy in the Brazilian electoral context. Together these papers establish browser extensions, data donations, and platform-independent experiments (championed by Allen2025-ot alongside the Meta2020 collaboration critiqued in Munger2025-cz) as the new infrastructure of platform accountability.
Detecting Coordination: From Co-Action Networks to Multiplex and Cross-Platform Models
Coordinated inauthentic behavior detection has matured into its own methodological subfield, and several papers here trace its internal arc. Yang2025-iv argues that purely time-based coordination signals are too easily evaded and proposes speed-and-frequency regularities as more robust. Iannucci2025-eg and Mannocci2025-ig independently converge on multiplex, time-aware network representations: aggregating modalities (retweets, mentions, hashtags, URLs) into a single flattened network discards structurally central nodes, while multiplex community detection preserves both monomodal findings and reveals new structures. Luceri2025-tr extends this enterprise to TikTok’s video-first ecosystem, where text-centric methods don’t transfer cleanly. Gerard2025-br reframes the problem entirely with CANE/t-CANE: instead of behavioral co-action, users are linked through shared participation in latent narrative clusters, enabling cross-platform analysis between Truth Social and X and surfacing a tiny set of “bridge users” carrying ~70% of migrating narratives. Pante2025-pq and Di-Marco2025-aa supply the necessary skeptical counterweights — re-examining inter-state coordination claims and showing that observed coordinated accounts on X exert far less cascade influence than commonly assumed, with placements statistically resembling random assignment. Minici2024-tf’s graph foundation model approach (IOHunter) gestures toward where this literature is heading next.
LLMs and Multimodal Methods as Both Tool and Threat
A cluster of papers wrestles with the rapid arrival of LLMs and VLLMs as substrates for social-science measurement. Achmann-Denkler2026-lx demonstrates GPT-4o outperforming specialized computer-vision pipelines for face recognition and person counting on Instagram campaign images, while Arminio2025-tw shows that VLLM-generated textual descriptions enable connotative (rather than merely denotative) clustering of climate-change imagery with interpretable TF-IDF summaries. DiGiuseppe2025-es pairs LLMs with paired-comparison designs to scale open-ended survey responses, and Sarmiento2025-as integrates LLMs into unsupervised framing analysis. But the tool-versus-threat duality is sharp: Lee2026-je shows GPT-4o can infer political alignment from ostensibly non-political Reddit and debate text — leveraging culturally politicized cues like “Tesla” or “Taylor Swift” — with serious privacy implications. Balluff2026-if supplies the field’s most pointed methodological reflection: LLMs are corporate products with opaque guardrails, environmental costs, demographic biases, and reproducibility problems, and for many tasks smaller encoder models or SVMs perform comparably. Fan2025-ut’s linear concept erasure (LEACE) on text embeddings is a more modest but exemplary case of principled methodological hygiene — purging source and language confounds from similarity measures rather than treating embeddings as neutral.
What Digital Traces Can and Cannot Tell Us
Even when data access is generous, several papers question the inferential leap from trace to population. Oswald2025-km foregrounds the production-consumption gap: a small minority of users produces most visible political content, distorting both citizen and researcher perceptions of public opinion. Green2025-ap cuts in a related direction, showing that domain-level audience-partisanship scores mistake heterogeneity for moderation — story-level “curation bubbles” reveal that moderate-scoring outlets serve very different partisan audiences for different articles. Luhring2025-od performs an analogous audit of NewsGuard, demonstrating that binary trustworthy/untrustworthy thresholds can flip empirical conclusions while continuous scores remain stable. Ulloa2024-jm interrogates measurement at the most basic level, showing that web-scraping ex-situ introduces ~34% content discrepancy relative to in-situ capture — and that paywalls and login walls bias scraped corpora systematically rather than randomly. Gaisbauer2025-by and Lai2024-to push for richer multi-level measurement of news circulation and ideological scaling, while Smith2025-kc uses a clever V1 API trick to identify the “tertius amplificans” — the causal mechanism by which influencer retweets generate transitive triads. Bruns2025-fz proposes “practice mapping” via vector embeddings as a way past the “hairball” limitations of conventional network visualization.
Networks, Migrations, and the Field’s Self-Understanding
Finally, several papers turn the methodological lens on the research community itself. Wang2026-ub tracks 7,542 academic early adopters attempting to migrate from Twitter to Mastodon, finding that even a coordinated, motivated subpopulation largely failed to sustain migration — with field-specific servers and federated engagement diversity the strongest retention factors. Scalco2026-bd operationalizes “information voids” as detectable imbalances between demand and supply across multiple data sources, exemplifying the kind of multi-platform triangulation that Yang2026-tq notes remains rare (82.59% of empirical papers still use a single platform). Taken together, these papers describe a field in flux: methodologically inventive, increasingly reflexive about its corporate entanglements (Heiss2026-qv, Bak-Coleman2026-mk, Munger2025-cz), and groping toward an infrastructure — regulatory (Peters2026-mo, Freelon2024-sc), technical (donations, browser extensions, multiplex models), and normative (Balluff2026-if) — adequate to studying platforms that no longer wish to be studied.