The infrastructure crisis: platforms as gatekeepers of knowledge
The backbone of this literature is a sustained argument that the tools of computational social science are hostage to corporate decisions about data access. Freelon2018-ao names the condition early, declaring a “post-API age” the moment Facebook’s Pages API closed in 2018, and Bruns2019-nr immediately reframes that closure not as privacy protection but as a strategic “APIcalypse” designed to frustrate critical, public-interest scrutiny. Freelon2024-sc and Giglietto2026-855a54cb extend this periodization across nearly two decades and six platforms, charting a move from laissez-faire APIs through a brief “Golden Age” of academic access to today’s pay-to-play or “walled garden” regimes — a trajectory Murtfeldt2025-wu quantifies bibliometrically, showing Twitter-based scholarship’s 25%-per-year growth cratering into decline after the $42,000/month Enterprise API. Yang2026-tq corroborates this from the demand side: social-media-based empirical research tripled through the early 2020s and then plateaued, with Twitter/X and Facebook remaining disproportionately dominant relative to their actual user bases, a mismatch between what is studied and what is used.
Against this closing door, the EU’s Digital Services Act emerges as the discipline’s main institutional hope and its own methodological problem. Giglietto2025-ed60bc90 surveys the resulting API/tool landscape, while Peters2026-mo shows that “data quality” — ostensibly a technical criterion — is in fact a politically contested object fought over by academics, NGOs, and platforms in the DSA’s Article 40 consultation process. Cullen2026-cb and Bruns2026-pn push this critique inward, documenting how CrowdTangle and its successor, the Meta Content Library, function as “epistemic infrastructures” and literal clean rooms that privilege code-fluent, well-resourced, Global North researchers while structurally excluding qualitative and cross-platform work. Rieder2025-ju and Zheng2026-bi supply parallel audits of YouTube’s and TikTok’s research APIs, finding systematic temporal decay, inconsistency, and non-representativeness that undermine even DSA-mandated systemic-risk research — motivating Zheng’s own open-source random-sampling tools (TubeStats/TokStats) as a corrective. Park2026-tr closes this cluster with a sobering reminder that the risks of working around these constraints are not merely technical: researchers doing scraping- or platform-adjacent work face real legal jeopardy, chilling effects, and uneven institutional support.
Detecting coordination: building the methodological toolkit
A second throughline concerns the specific problem of coordinated inauthentic behavior, tracing an arc from conceptual groundwork to tool-building to formal statistical critique. Giglietto2019-e9be81c1 lays philosophical groundwork, arguing that “fake news” research should shift attention from creators’ intent to the propagation dynamics that actually drive viral cascades — a reframing that anticipates later network-based coordination detection. Giglietto2022-b30e8b4e then turns a methodological lens on a widely used dataset itself, showing that Meta’s 100-public-shares threshold in the URL Shares Dataset introduces a platform-wide algorithmic artifact (likely tied to Feed-ranking changes) that can masquerade as cross-country social signal — a cautionary tale for anyone using derived platform datasets for longitudinal or comparative coordination research. Yang2025-iv directly advances detection methodology, proposing that the speed and frequency regularities of link-sharing are a more evasion-resistant coordination signal than the post-timing features exploited by earlier tools, while Luceri2025-tr extends the broader network-similarity paradigm to TikTok’s video-first ecosystem, arguing that text-centric CIB detection frameworks do not transfer cleanly to new modalities. Iannucci2026-my caps the cluster with a foundational critique: reframing coordination detection as a planted-submatrix hypothesis-testing problem, it shows that the choice of null model is decisive for what counts as “detectable,” that density of a recovered cluster is not itself evidence of coordination, and that combining modalities yields principled (if sub-linear) gains — a corrective lens that applies retroactively to the whole family of coordination-detection tools surveyed here.
Recruitment, sampling, and the validity of the data we collect
A parallel set of papers interrogates not detection algorithms but the basic act of getting representative or usable data in the first place. Iannelli2018-ebd918b7 pioneers Facebook-ads-based recruitment for niche, hard-to-reach populations (Italian conspiracy believers), using Pixel tracking and custom-audience exclusion to produce a genuine conversion-rate metric — efficient and cheap, though its effectiveness at isolating ideologically distinctive respondents remains unproven. Stagnaro2025-pz broadens the lens to nine opt-in online panels more generally, showing that representativeness and response validity trade off against each other and that simple attention checks can improve data quality cheaply — directly relevant to anyone choosing a panel for digital-trace-adjacent survey work. Ulloa2024-jm turns to web-tracking methodology itself, demonstrating that the collection environment (in-situ browser capture vs. ex-situ scraping), not time delay, is the dominant and systematically biased source of error in scraped news-consumption data — a finding with direct implications for any study using post-hoc scraped content as a proxy for exposure. Oswald2025-km rounds out this group conceptually, arguing that the “production-consumption gap” — a small minority producing most visible content — distorts not just citizens’ perceptions of public opinion but researchers’ inferences from platform trace data, an implicit warning to every study in this bibliography that treats visible posting behavior as representative of a public.
New computational instruments and reflexive critique of LLMs
A third cluster develops or critically evaluates emerging computational instruments. Arminio2025-tw and Fan2025-ut both push embedding-based methods past their naive baselines: Arminio shows that VLLM-generated textual descriptions of images yield more interpretable, “connotative” clustering than CNN pipelines, while Fan demonstrates that linear concept erasure (LEACE) can remove source/language confounders from text embeddings, substantially improving clustering and retrieval without degrading out-of-domain performance. Le-Mens2025-qz proposes a complementary LLM application — prompting and averaging to scale political texts along ideological dimensions — extending the Wordscores/Wordfish tradition into the LLM era. Bruns2025-fz offers a structural innovation for relational data, proposing “practice mapping” via vector embeddings of network actions to escape the “hairball” problem of conventional multimodal network visualization. Fan2026-af reorients the field temporally, arguing that digital trace data’s fine-grained timestamps should be mined as “user-sequences” via six computational approaches (sequence analysis, HMMs, process mining, embeddings, etc.) rather than aggregated away. Running against this wave of enthusiasm, Balluff2026-if supplies the necessary counterweight: a critical audit of LLM adoption in communication research showing that synthetic data, LLM-based text analysis, and LLM simulations carry underacknowledged reproducibility, bias, environmental, and ethical costs, and arguing for a “least resource-intensive adequate method” norm — a methodological corrective that should be read alongside Arminio’s and Le-Mens’s more optimistic proposals.
Substantive studies as methodological demonstrations
Finally, several empirical studies function less as topical contributions than as showcases — and stress tests — of the methods above. Barbera2015-je and Barbera2015-fw remain foundational demonstrations of network-based ideal-point estimation from follower graphs, establishing that homophilic following behavior can proxy for ideology and that polarization is topic- and time-contingent rather than uniform — a measurement strategy whose assumptions later papers (e.g., on Reddit) explicitly complicate. Dehghan2026-sy applies multi-indicator overlap analysis (cross-posting, cross-commenting, link-sharing) to show that polarization on political Reddit is “sedimented” through an entanglement of platform design, moderation, and discourse rather than reducible to any single measurable signal — a finding that cautions against single-metric polarization measures of the Barberá type. Graham2026-fb and Wang2026-ub both use mixed-methods platform ethnography (of “botting” accounts and of academic Twitter-to-Mastodon migration, respectively) to destabilize clean technical categories — human/bot, successful/failed migration — that quantitative pipelines often take for granted. Gaisbauer2025-xx—rather, Gaisbauer2025-pz is not the key; the paper Gaisbauer2025-xx placeholder aside—Gaisbauer2025-pz