The arc of inquiry
The papers filed here trace a rapid arc from documenting a specific case of AI-enabled influence — gambling promotion imagery flooding Facebook — outward to a much broader reckoning with what generative AI does to disinformation, persuasion, and the epistemic infrastructure of public communication. Read together, they move from empirical detection (what is actually happening on platforms) through experimental measurement (how persuasive and how novel is this, really) to theoretical and pedagogical reframing (what new concepts and what new researchers do we need to make sense of it).
Coordinated influence operations go synthetic
The anchor case is Giglietto2026-9b6a992d, which traces the visual persuasion machinery of coordinated gambling-promotion networks on Facebook and documents an exponential, structurally-broken surge in posting volume that tracks almost exactly with ChatGPT’s public launch. Its central argument — that generative AI does not invent new persuasion drivers but intensifies and recombines existing ones (aspirational wealth, manufactured trust, FOMO, cultural localization) at industrial scale — becomes a template picked up elsewhere in the collection. Giada2026-fc9a3833 extends this logic by placing the gambling case alongside a state-aligned pro-Putin network and an unmoderated adult-content network, arguing that all three share a common operational grammar of concealment, affordance exploitation, and coordination that existing “coordinated inauthentic behaviour” and false-news taxonomies cannot capture — a structural, rather than motivational, definition of manipulation that generative AI slots neatly into. Luceri2025-tr pushes the empirical frontier toward video-first platforms, showing that TikTok’s architecture demands new detection methods rather than a straightforward port of text-centric CIB techniques, while Orlando2025-ul simulates the underlying mechanism computationally: LLM agents given nothing more than mutual awareness of teammates spontaneously converge on the coordination signatures (narrative homogenization, synchronized amplification, rapid hashtag adoption) that real operations exhibit, suggesting that “coordination” may increasingly be an emergent property of agentic systems rather than a centrally scripted campaign. On the political-communication side, Perez-Curiel2026-ld and Nangle2026-yo document how far-right actors weave AI-generated multimedia and algorithmically-favored affective softening (humor, memes, positive affect) into mainstreaming strategies on Instagram and X, complementing Schroeder2026-im’s more speculative but structurally grounded warning that LLM-agent swarms could soon manufacture synthetic consensus, poison training data (“LLM grooming”), and erode institutional trust at a scale beyond anything IRA-style botnets achieved.
How persuasive is it, really?
A second cluster interrogates the actual causal power of generative AI, and the results complicate the alarm implicit in the operations literature. Hackenburg2026-ud delivers the most striking finding: frontier conversational AI out-persuades even world-champion debaters and professional canvassers, an advantage rooted not in charisma or personalization but in sheer information throughput (fact density), and one that only disappears when AI is artificially slowed to human typing speed. Lin2025-xp corroborates this across four real elections, finding brief AI dialogues shift votes more than traditional ads, again mainly through evidence rather than manipulation — though with a troubling asymmetry, as models advocating right-leaning candidates are systematically less accurate. Kotz2026-lk shows the same evidence-based dialogic mechanism can be turned toward prosocial ends, moving skeptics on climate, vaccination, and inequality policy more than the already-convinced. Against this, Hameleers2026-mc offers a useful corrective specifically to the “visual disinformation” strand: AI-generated images do not universally outperform text or even decontextualized real video, and the panic over synthetic imagery’s unique power may be overstated relative to older, low-tech tactics. Wack2026-bt complicates the picture further by showing that deepfake impact in a real electoral context (Kenya) is not fixed at the point of generation at all, but socially negotiated through the comments and interpretive cues surrounding a clip — a “collective sensemaking” account that sits in productive tension with individual-level persuasion experiments like Hackenburg’s and Lin’s.
Epistemic infrastructure under strain
A third thread asks what happens to the concept of truth itself once synthetic content becomes ambient. Emilio2026-ik offers the collection’s most totalizing framework, arguing the real danger is not discrete deepfakes but layered “synthetic realities” (content, identity, interaction, institutions) that produce a paradox: as fabrication becomes cheap and ubiquitous, rational actors discount all digital evidence, taxing verification everywhere. Dierickx2026-tw approaches the same problem from fact-checking epistemology, proposing “emergent facts” as a category distinct from evidence-based, interpretive, or rule-based facts — explicitly framed as a teaching-relevant lens for cultivating AI literacy among the next generation of journalists and researchers. Waight2026-ts locates a subtler erosion further upstream, showing that state-controlled media is already baked into LLM training corpora and measurably skews model outputs toward pro-regime framings by query language — disinformation laundered not through generation but through training data. Triedman2025-uy’s forensic comparison of Grokipedia against Wikipedia is a vivid case study of this same dynamic institutionalized: an AI encyclopedia that is derivative of crowdsourced knowledge yet systematically cites lower-quality and ideologically aligned sources on contested topics, illustrating how “objective-seeming” AI infrastructure can quietly reshape the reference layer of public knowledge. Mattis2026-gu turns to audience-side epistemics, finding that any disclosure of AI involvement in journalism dents trust, with fact-checking tasks penalized most — a finding with direct implications for how newsrooms, and by extension journalism educators, should think about transparency.
Reframing the field: from social media to post-social, from slop to attunement
The remaining papers are less about specific harms than about the conceptual vocabulary future researchers will need. Tornberg2026-lc makes the boldest structural claim: the “social media” paradigm itself is dissolving into algorithmic broadcasting, semi-private micro-communities, and AI-mediated communication, with generative AI a central solvent severing platforms from human-generated content. Baym2026-tr revisits an older diagnosis of platform harms a decade on and finds generative AI has deepened rather than displaced them — feeding on user data, enabling non-consensual imagery, and consolidating wealth — while calling for researchers to move past critique toward engaged intervention, a normative note that resonates with the pedagogical undertone of this whole topic. Nguyen2026-vm examines how news media themselves frame LLMs, finding mentalistic-agentic language is real but neither dominant nor monolithic, varying by newsroom rather than region — a caution against overreading anthropomorphism as uniform hype, and a reminder that critical AI literacy must be taught with empirical nuance rather than received wisdom. Hollingshead2026-vx shows platform vernacular (TikTok affordances) is shared symmetrically across ideological camps even as it enables both inclusion and xenophobia, with generative AI beginning to appear as a tool for “speculative worldbuilding” in anti-immigrant content. Finally, a trio of cultural-theoretical pieces — Galip2026-ix, Tang2026-gu, and Topinka2026-hb — supply the vocabulary for the aesthetic residue of this whole system: “slop” and “brainrot” as gimmicks that work simultaneously too little and too hard; Chinese creators industrializing virality itself into a manageable production process; and “slopaganda” as content that persuades not by deceiving but by affectively attuning itself, via alignment and RLHF, to the moods of the feed — intensifying existing ideological formations rather than installing new beliefs.
Political economy and governance
Bracketing the empirical and theoretical work are two papers attentive to regulation and comparative political context. Schiffrin_undated-gi maps the “scam ecosystem” of deepfake financial fraud and argues persuasively for shifting liability from victims to platform and financial gatekeepers, an argument that dovetails with Giglietto and Giada’s diagnosis that amplification infrastructure, not just content moderation, is the proper regulatory target. Beacken2026-zb provides the necessary comparative counterweight to any universalizing narrative, showing across six democratically weakened states that GenAI propaganda adoption is shaped by local political, economic, and social structure rather than following one deterministic trajectory — a fitting closing note for a topic that, across every paper here, insists that generative AI’s disinformation effects are neither uniform nor inevitable, but are actively co-produced by platforms, institutions, audiences, and the researchers now being trained to study them.