From open APIs to walled gardens: the collapse of a research infrastructure
The starting point for almost everything else in this collection is a historical rupture. Freelon2018-ao named the moment—Facebook’s 2018 closure of the Pages API—that ended the era in which computational social scientists could treat platform APIs as stable, TOS-compliant infrastructure, and urged researchers to build more durable skills (scraping) alongside a sharper sense of the legal and ethical stakes of doing so. Freelon2024-sc extends this into a full periodization, tracing five eras of access from “laissez-faire” to today’s fractured landscape of pay-to-play APIs, academic walled gardens, and unofficial workarounds, and shows that platforms facing the most scrutiny (Meta, X) have become the most closed. Bastos2025-ya supplies the elegy for the platform that made this infrastructure visible in the first place, reviewing how Twitter’s suite of APIs underwrote an entire subfield of “Twitter studies” now largely orphaned by X’s pricing regime. Murtfeldt2025-wu and Yang2026-tq convert this narrative into bibliometric evidence: both document explosive growth in social-media-based research through the early 2020s followed by stagnation or decline once free API access ended, with Yang2026-tq additionally showing that research attention has always been skewed toward Twitter/X and Facebook independent of where actual public attention lives, and that scholars are only belatedly diversifying toward YouTube, TikTok, and donated data.
What has replaced open APIs is not simply less access but differently governed access, and several papers interrogate that governance directly. Peters2026-mo shows that “data quality” under the EU’s Digital Services Act is less a technical property than a politically contested category, actively pushed onto the regulatory agenda by academic and NGO stakeholders over platform resistance. Bruns2026-pn and Cullen2026-cb examine the resulting “clean room” model empirically: Bruns2026-pn documents how the Meta Content Library’s Jupyter-only, data-deleting, code-centric design structurally excludes qualitative and cross-platform research and widens Global North/South inequities, while Cullen2026-cb’s interviews with CrowdTangle users develop a “seeing like an API” framework showing how a tool’s technical design and governance rules quietly shape what questions researchers can even conceive of asking. Rieder2025-ju and Zheng2026-bi push this critique to platforms often treated as comparatively open: Rieder2025-ju’s longitudinal audit finds YouTube’s search API “forgetful by design,” with findable results decaying sharply within weeks of publication in ways that directly undermine DSA-style systemic-risk research; Zheng2026-bi responds constructively by building genuinely random-sampling tools (TubeStats, TokStats) to supply the “denominators” that keyword- and popularity-based samples cannot.
Bias hiding in the data you’re given
A second, closely related line of argument shows that even where platforms do provide research data, that data systematically misrepresents the phenomena it claims to describe. Giglietto2022-b30e8b4e and Allen2021-ai both dissect the same artifact—the 100-public-shares threshold in Meta’s Social Science One URL dataset—and reach complementary conclusions: Giglietto2022-b30e8b4e shows that changes in Feed ranking (not real-world events) produce eerily similar share trajectories across 46 countries, while Allen2021-ai demonstrates that the threshold inflates apparent fake-news prevalence by roughly 4x relative to a representative panel, because virality-optimized and publicly-shared content crosses the cutoff disproportionately. Ulloa2024-jm identifies an analogous distortion in web-tracking research: ex-situ scraping (vs. in-situ capture) introduces far more content disparity than the time delays methodologists have worried about, and the errors are non-randomly distributed across news categories, biasing conclusions about media diets. Luhring2025-od audits a different infrastructure entirely—the NewsGuard trustworthiness database widely used in misinformation research—finding its continuous scores fairly stable since 2022 but its binary trustworthy/untrustworthy cutoff highly sensitive to a handful of borderline sources crossing threshold, a methodological echo of the Facebook share-threshold problem. Anwar2024-34dba628 and Bastos2025-ol show how even seemingly simple platform signals—Facebook Reactions, visual self-presentation in troll-farm profile images—carry systematic biases and interpretive ambiguities that complicate their use as proxies for sentiment, controversy, or attribution. Oswald2025-km frames the deepest version of this problem: because a small minority of highly active users produce most visible content, both citizens and researchers who read platform data as a mirror of “public opinion” are seeing only the tip of an iceberg. Wan2026-ai illustrates how much is nonetheless salvageable through careful digital-trace design, using donated Google Discover histories to study political-efficacy-conditioned information selection within algorithmic feeds—a reminder that trace-data problems are not uniform but depend on what is being measured and how.
Who controls the evidence: industry influence and legal exposure
A third cluster interrogates the political economy behind data access rather than its technical artifacts. Heiss2026-qv and Bak-Coleman2025-pm both argue, via explicit analogy to tobacco, pharma, and food-industry science, that platforms’ monopoly over the data needed to study them creates a uniquely acute conflict-of-interest problem—one that open-science practices like preregistration cannot neutralize, because platforms still control what questions are askable and what results are visible (as in Meta’s contested 2020 election studies). Bak-Coleman2026-mk supplies the empirical backbone for this argument, finding that roughly half of high-profile social media papers in Science, Nature, and PNAS carry disclosable industry ties—most undisclosed—concentrated among a small set of repeat authors, editors, and reviewers, with industry-tied work skewed toward user-blaming topics like misinformation sharing and away from platform-dynamics research, and receiving double the citations and public attention. Park2026-tr complements this institutional critique from below: its interviews with researchers facing legal threats under laws like the CFAA and Computer Misuse Act document real chilling effects—abandoned projects, vulnerability stockpiling, career damage—showing that the precarity of platform-critical research is not only a matter of API paywalls but of researchers’ personal legal exposure when the only route to needed data is unauthorized scraping or terms-of-service violation, directly recalling the tension Freelon2018-ao first flagged.
Alternatives to platform data: ads, panels, and the return of the survey
Against this backdrop of restricted and distorted platform data, two papers examine older or adjacent methodologies for reaching populations directly. Iannelli2018-ebd918b7 develops a Facebook-ad-based recruitment procedure using Pixel tracking and custom-audience exclusion to reach a hard-to-survey population (conspiracy-theory sympathizers) cheaply and controllably, though its comparison against a general-population benchmark leaves open whether ad-based “interest” targeting actually identifies ideologically distinctive respondents. Stagnaro2025-pz steps back to systematically compare nine opt-in online panels (Prolific, Lucid, MTurk, and others) on response validity, representativeness, and professionalism, finding a real trade-off between demographic-quota-driven representativeness and response quality, and showing that two simple attention checks can substantially improve validity without much cost to representativeness—useful groundwork for anyone choosing a sample-recruitment strategy in a post-API landscape where digital trace data is harder to get.
The LLM moment: efficiency, bias, and reflexive adoption
A fourth thread turns to the newest disruptor of research method: large language models. Balluff2026-if offers the field’s most pointed caution, arguing that LLM adoption in communication research for text analysis, synthetic data generation, and simulation has outpaced critical reflection on reproducibility (opaque corporate updates), demographic and linguistic bias, environmental cost, and the validity risks of prompt sensitivity—counseling a “least resource-intensive method” ethic and preference for open models. Brown2025-jk provides more reassuring, fine-grained evidence on one specific worry—demographic bias in LLM annotation—showing across four contentious datasets that biases are dataset-specific rather than model-specific, small in magnitude, and dwarfed by item difficulty (human label entropy) as a predictor of LLM-human agreement. Two papers then show LLMs put to constructive methodological use: Ober2026-vd integrates topic modeling with LLM-assisted labeling in a human-in-the-loop qualitative coding workflow for interview transcripts, arguing this preserves interpretive transparency better than end-to-end neural approaches; Arminio2025-tw shows that Vision-LLMs, by generating connotative textual descriptions of images before clustering, substantially outperform CNN-based pipelines at capturing culturally embedded meaning (e.g., in climate-change Instagram imagery) while remaining interpretable via keyword summaries—directly answering Balluff2026-if’s call for well-justified, task-appropriate tool choice.
Innovating on—and complicating—the analytic toolkit
A final set of papers pushes on methodological practice for handling the data researchers do manage to obtain, and on the interpretive instability that results from unreflective method choice. Bruns2025-fz proposes “practice mapping,” using vector embeddings of network actions to escape the interpretive limits of the network-visualization “hairball.” Fan2026-af argues communication research has under-used the temporal richness of digital trace data, comparing six computational approaches (sequence analysis, HMMs, process mining, embeddings) on cross-platform donated data to show that no single method captures all facets of behavioral dynamics. Hartmann2025-px and Schemer2026-mh both demonstrate, at the level of literature and of a single dataset respectively, how much apparent empirical disagreement is actually a methodological artifact: Hartmann2025-px’s review of 129 echo-chamber studies attributes contradictory findings largely to inconsistent conceptualization and operationalization, while Schemer2026-mh’s specification-curve analysis of 504 measurement combinations for partisan-media-slant and polarization shows that method choice swings effect magnitudes sevenfold, even though it never flips their direction—an implicit vindication of Hartmann2025-px’s diagnosis. Two earlier landmark studies, Bail2018-fk and Barbera2015-fw, stand as touchstones for what rigorous platform-dependent computational social science looked like before the access crisis—showing, respectively, that cross-partisan exposure can backfire rather than depolarize, and that ideological echo chambers are topic- and time-contingent rather than blanket features of Twitter discourse—and both now face the harder question of whether their designs remain replicable at all under current data regimes. Wang2026-ub offers a natural experiment in this destabilized landscape, tracking academics’ failed migration from Twitter to Mastodon and finding that even highly motivated, well-connected early adopters mostly drifted back, with retention driven by discipline-specific servers and cross-server engagement diversity rather than network size—a case study in how platform shocks reshape not just data access but the very communities researchers study.
Coda: a qualitative counterpoint
Not all inquiry into platform culture depends on API access or computational scale. Volpe2026-um’s digital ethnography of Italian micro-influencers—built from observation and interviews rather than trace data—stands as a reminder that questions about platform pressure, authenticity, and labor can be pursued through methods largely immune to the access crises dominating the rest of this collection, even as its subjects (creators optimizing for algorithmic visibility) are shaped by exactly the platform infrastructures whose data researchers are increasingly locked out of.