The arc: from platform studies to a generative-AI-mediated epistemic crisis
This cluster traces a single, increasingly urgent argument: that generative AI is not merely adding a new content type to the existing media ecosystem but is restructuring the ecosystem’s basic logics of persuasion, coordination, and truth. The papers move from macro-theoretical reframings of what “media” even is now, through tightly controlled experiments on AI’s persuasive mechanics, into empirical case studies of synthetic imagery deployed for commercial and political ends, and finally into proposals—both epistemological and regulatory—for how societies might counter what several authors call an emerging “epistemic crisis.”
Reframing the terrain: post-platform, post-social
Three papers set the conceptual stage by arguing that the categories inherited from a decade of platform studies no longer fit. Tornberg2026-lc diagnoses the unraveling of the “social media” paradigm itself, pointing to algorithmic broadcasting, semi-private micro-communities, and AI-mediated communication as three successor formations—crucially, one of these formations (chatbot interaction) involves no circulating “content” or public at all, severing the link between platforms and human-authored discourse. Helmond2026-ll revisits “platformization” a decade after coining the term, arguing that platform logics are now being reorganized around AI infrastructure (inference, recommendation, compute) rather than behavioral-signal harvesting from the open web—Big Tech is becoming “Big AI.” Tornberg2026-im synthesizes this shift most systematically, proposing that the “platform society” is giving way to an “AI society” organized around cognitive lock-in (economic), generative rather than predictive epistemics (epistemic), and alignment politics displacing visibility politics (political). Together these three essays supply the theoretical vocabulary—synthetic content without users, platform-readiness, cognitive enclosure—that the more empirical papers in this set implicitly presuppose.
The mechanics of AI persuasion
A second cluster asks, empirically, why conversational AI persuades, stripping away speculation in favor of large-scale experiments. Hackenburg2025-dj isolates information density—the sheer volume of fact-checkable claims—as the dominant lever of persuasiveness, more powerful than model scale or personalization, but shows this comes bundled with a persuasion–accuracy trade-off. Hackenburg2026-ud escalates the comparison to elite human persuaders (championship debaters, professional canvassers) and finds AI systematically outperforms them, again implicating information throughput rather than rapport or personalization, with effects extending to real monetary donations. Lin2025-xp replicates these findings across four real elections, showing AI dialogue shifts voting intentions more than traditional ads, driven by evidence-based argumentation rather than manipulation tactics—while noting a troubling partisan asymmetry in the accuracy of AI claims favoring different political sides. Read together, these three papers push back against alarmist “microtargeting” narratives while establishing an arguably more consequential finding: AI’s persuasive power is scalable, structural, and tied to a mechanism (fact density) that is largely orthogonal to truthfulness.
Synthetic imagery as coordinated persuasion: the gambling and political cases
The empirical heart of the collection examines generative AI’s use in actual coordinated campaigns. Giglietto2026-9b6a992d is the paradigm case named in this topic’s framing: an analysis of over 2,300 images from coordinated Facebook gambling networks showing an exponential, structurally-broken rise in synthetic visual content after ChatGPT’s launch, with AI not replacing but intensifying existing persuasion drivers (aspirational wealth, manufactured trust, FOMO) while exploiting the regulatory grey zone between paid and organic content. This logic of AI-intensified, platform-exploiting visual persuasion recurs in explicitly political contexts: Perez-Curiel2026-ld documents far-right adoption of AI-generated multimedia in the 2024 European elections; Crichton2026-kn shows experimentally that commercial image models, when fed far-right extremist imagery, sanitize explicit markers of hate while preserving or displacing the underlying ideology—evidence that the affinity between generative AI and far-right visual culture is structural, not accidental; and Topinka2026-hb theorizes this material as “slopaganda,” arguing its power lies not in deceiving viewers (as with deepfakes) but in affectively “attuning” to feed-native moods, intensifying rather than disrupting existing ideological formations. Galip2026-ix and Tang2026-gu extend this into the broader cultural economy of “slop”: the former theorizes slop/brainrot via Ngai’s gimmick as simultaneously laborless and over-laboring content that exposes contradictions of platform capitalism, while the latter grounds this in Chinese creative labor practices where virality itself is industrially engineered rather than organic—showing that synthetic persuasion is as much a production-labor phenomenon as a content phenomenon.
Platform-specific vectors and multi-agent coordination
Several papers examine how specific platform affordances channel synthetic persuasion. Bosch2024-hj identifies sound as an underexamined, TikTok-specific propaganda vector in Ukraine-war disinformation; Hollingshead2026-vx complicates assumptions about platform hostility by showing Canadian TikTok immigration discourse skews pro-immigration, while flagging generative AI’s emergent use for racist “speculative worldbuilding.” At the level of coordination mechanics rather than content, Orlando2025-ul and Schroeder2026-im model how LLM-driven agents themselves generate coordination: the former demonstrates in simulation that even minimal “teammate awareness” among LLM agents produces coordination nearly equivalent to explicit collective deliberation, while the latter extrapolates this into a policy-facing warning about “malicious AI swarms” capable of manufacturing consensus, poisoning training data, and eroding institutional trust—pairing the technical finding with governance proposals (provenance tools, an “AI Influence Observatory,” incentive-based disruption of manipulation markets).
Deception, fraud, and the social negotiation of belief
A related strand interrogates how audiences actually process synthetic deception, pushing against simple deception-based models. Hameleers2026-mc finds experimentally that AI-generated visual disinformation does not universally outperform text, and that low-tech video decontextualization can be just as persuasive as high-tech synthesis—tempering alarm about generative AI’s uniqueness. Wack2026-bt extends this into a Kenyan electoral context, showing deepfake impact is not fixed at production but “socially negotiated” through surrounding audience commentary, moderated heavily by partisanship. Schiffrin_undated-gi shifts from political belief to financial harm, mapping the “scam ecosystem” of deepfake-enabled fraud and arguing regulation must target platform and institutional gatekeepers rather than individual victims. Across these three, the common thread is a rejection of purely cognitive, decontextualized accounts of deepfake harm in favor of socio-technical and infrastructural ones.
Epistemology and governance: naming and countering the crisis
Finally, two theoretical papers attempt to name the epistemic stakes these empirical findings imply. Dierickx2026-tw argues that traditional fact-checking categories (evidence-based, interpretative, rule-based) cannot capture generative AI outputs, proposing “emergent facts” as a new, relationally-evaluated category. Emilio2026-ik goes further, arguing the real danger is not isolated deepfakes but coherent “synthetic realities” spanning content, identity, interaction, and institutions, culminating in a paradox whereby ubiquitous synthetic media causes societies to rationally discount all digital evidence, raising “the cost of truth itself.” Beacken2026-zb grounds these abstract risks comparatively, showing across six democratically weakened countries that GenAI-propaganda adoption is not a uniform technological inevitability but is mediated by local political, economic, and social structures—a necessary corrective to the more deterministic narratives found elsewhere in the set, and a fitting close to the arc: generative synthetic manipulation is real and structurally embedded, but its uptake, form, and consequences remain stubbornly contextual.