The imaginative landscape of AI: Locating Silicon Valley’s “quiet futuring”

Summary

Despite the title metadata, this paper is an infrastructure-oriented case study of the 2025 German federal election that examines the practical, methodological, and governance bottlenecks of monitoring electoral discourse across X and TikTok. Collecting 81,866 X posts and 43,597 TikTok videos, the authors document how platform-specific affordances, restrictive APIs, and divergent party strategies make reproducible multi-platform analysis extremely difficult under current conditions. They argue that “platform observability” should be reframed as a shared, reusable research infrastructure — standardized tools, workflows, and documentation — rather than as a series of one-off empirical studies, and they derive four recommendations to that end.

Key Contributions

  • An empirical multi-platform, multi-modal dataset spanning X and TikTok across seven major German parties during the 2025 federal election campaign.
  • Concrete documentation of post-API bottlenecks: rate limits, algorithmically filtered results, noisy hashtag-based sampling, and obstructed access.
  • A reframing of election observability as a research infrastructure problem rather than a study-by-study data-collection problem.
  • Four actionable recommendations: standardize platform-specific data structures; raise data-access quality standards (including transparent labeling of political accounts); tailor analytical methods to content type (multimodal for video); and institutionalize interdisciplinary computational/political-science collaboration.
  • Empirical input for ongoing DSA Article 40 vetted-researcher-access debates.

Methods

  • Case study window: 6 November 2024 – 23 February 2025.
  • Parallel scraping of X (X API v2 via a twitter-api-client implementation) and TikTok, seeded with official party accounts and expanded via mentions (X) and hashtags (TikTok).
  • Distinction between authored (party-account) and promoted (retweets, mentions, shared hashtags) content.
  • Multimodal TikTok pipeline: pyannote.audio (speaker diarization), OpenCV + DeepFace (facial emotion), HuBERT (vocal tone), XLM-R (textual sentiment); Cramér’s V for cross-modal association.
  • BERTopic for topic modeling and cross-platform/temporal comparison.

Findings

  • Dataset: 81,866 X posts from 541 accounts; 43,597 TikTok videos from 4,682 accounts.
  • Strong cross-party divergence in platform preference: BSW dominates on X with minimal TikTok presence; AfD is highly active on TikTok; FDP and CDU active on both.
  • Cross-modal associations on TikTok are uniformly weak (Cramér’s V ≤ 0.104) — text, voice, and face carry largely independent emotional signals.
  • Political TikToks frequently combine negative textual sentiment and expressive faces with calm vocal delivery, consistent with a “politainment” register.
  • Topic overlap exists (Russia/Ukraine, migration dominate both), but migration is more prominent on TikTok, while EU and Israel/Gaza skew toward X; acute news events (Magdeburg attack, Syrian regime fall) surface on X but not in TikTok top topics.
  • ~2/3 of hashtag-sampled TikTok content is non-political noise needing aggressive filtering.
  • X API v2 caps (~1000 posts/day per account) preclude full historical recovery; TikTok returns partial, algorithmically shaped results.

Connections

This paper sits squarely in the post-APIcalypse debate over what computational election research can still do, sharing concerns with work on data access constraints and platform-governance critique such as Rieder2025-ju, Rieder2026-pp, and Ohme2026-nv. Its push to treat observability as durable shared infrastructure connects to the social media observatory and monitoring-workflow agenda exemplified by Bruns2025-fz, Bruns2026-yv, and Helmond2026-ll, while its multi-platform, multimodal TikTok analysis of electoral discourse resonates with platform-comparative election work such as Bouchaud2026-lr and Achmann-Denkler2026-lx.

Podcast

A research-radio episode discusses this paper: Listen