Giglietto, F., Terenzi, M., Chakraborty, A., & Marino, G. (2026). Synthetic seduction: Evolving visual persuasion in coordinated online gambling promotion with generative {AI}. Countering Disinformation in the Era of Generative AI. https://doi.org/10.1007/978-3-032-11782-3_4
Summary
This paper investigates how generative AI has reshaped coordinated organic gambling promotion on Facebook. Analyzing 2,323 images from 223 coordinated public groups (surfaced via the Vera AI alerts workflow), the authors build a mixed-methods pipeline — VLM-generated denotative/connotative descriptions, embeddings, UMAP+HDBSCAN clustering, and qualitative co-occurrence coding — to derive a typology of visual persuasion drivers. They argue that generative AI has not replaced established persuasion strategies (aspirational wealth, manufactured trust, FOMO, gamification, celebrity endorsement, cultural localization, social-relations exploitation) but intensified and recombined them into hyper-real, immersive imagery deployed at industrial scale. A structural break in July 2023, following ChatGPT’s launch, marks an exponential jump in posting volume, exploiting Meta’s asymmetric governance that regulates paid but not organic gambling content.
Key Contributions
- An empirically grounded typology of visual persuasion drivers in coordinated organic gambling promotion.
- Evidence that generative AI accelerates and intensifies — rather than replaces — existing persuasive architectures.
- A reproducible mixed-methods pipeline combining VLM-based image description, dual denotative/connotative embeddings, density-based clustering, and human qualitative coding.
- Identification of Meta’s paid-vs-organic regulatory asymmetry and algorithmic amplification as central governance blind spots.
- Reflexive discussion of using LLMs both as analytic instruments and as the very technologies producing the studied manipulative content.
Methods
Coordinated groups (n=223) were identified through Vera AI’s coordinated link-sharing detection (14-second window, 0.995 edge weight), seeded from accounts spreading fact-checker-flagged material. Posts and images (10,671 posts; 2,323 images; 2017–2024) were collected via the Meta Content Library and a custom image downloader. GPT-4o produced structured denotative and connotative descriptions, embedded with text-embedding-3-small, reduced with UMAP, and clustered with HDBSCAN (101 denotative, 51 connotative clusters). A 366-cell co-occurrence matrix guided qualitative coding by four coders of 85 combinations to thematic saturation. Temporal effects of ChatGPT’s launch were tested via t-tests, Wilcoxon tests, interaction-term regression, and structural break detection.
Findings
- Aspirational wealth/hyper-masculine motifs appeared in ~55% of analyzed cluster combinations; transactional “trust proof” visuals (payment receipts, cash-out screenshots) in ~37%.
- Drivers include aspirational wealth, manufactured trust, FOMO/urgency, gamification, celebrity endorsements (e.g., Manny Pacquiao), social-relations exploitation, and cultural localization (Filipino, Urdu).
- An Urdu-language cluster embeds gambling within emotionally charged, morally conservative family narratives — an ideologically inflected localization strategy.
- Mean monthly posts rose from 2,121 (pre-ChatGPT) to 280,952 (post-ChatGPT), a 13,242% increase; regression confirmed level and slope shifts (p<0.0001), with a structural break in July 2023.
- Post-2022 imagery shows consistent AI-generation markers (hyper-real lighting, smoothed surfaces, dreamlike saturation, improbable juxtapositions such as sharks with slot machines) and stacks multiple drivers per image.
- Two emblematic AI-generated posts reached 4.3M and 3.3M views, propagating across coordinated groups.
Connections
This paper sits at the intersection of coordinated inauthentic behavior research and generative-AI-enabled influence, extending the authors’ prior work on coordinated link-sharing detection (Giglietto2020-9d8acdd7, Giglietto2022-0e951ac5, Giglietto2023-fa71a001, Giglietto2024-cbeb3f70, Marino2024-2fbc690f) into the visual and synthetic-media domain. It complements work on AI-generated imagery in political and manipulative contexts (Achmann-Denkler2026-lx, Dodds2026-df, Stanusch2026-ec) and on AI-driven or automated coordinated operations (DiGiuseppe2026-pu, DiGiuseppe2025-es, Minici2024-tf, Luceri2025-tr, Ng2026-og). Its emphasis on algorithmic amplification as a governance blind spot resonates with platform-power critiques such as Gillespie2026-aa.
Podcast
A research-radio episode discusses this paper: Listen