Generative AI and Media
From discourse to infrastructure: locating AI in the media ecosystem
A first cluster of papers treats generative AI primarily as a discursive object whose meaning is being actively negotiated in and through news media. Nguyen2026-vm shows that mentalistic-agentic framings of LLMs coexist with technical and adversarial registers, and that anthropomorphism is not monolithic but does varied rhetorical work — uncritical hype, routine explanation, and pointed critique. Wang2025-zy complements this by mapping utopian/dystopian imaginaries across UK, US, Chinese, and Indian press, while Stanusch2026-ec zooms in on a critical moment — the Altman firing — to show how industry actors deploy “premediation” and “preclusion” to monopolise the future of AI. Dodds2026-df reframes these patterns infrastructurally: AI hype is not a bubble but a structuring force that redistributes legitimacy, sourcing, and resources within journalism itself. Weinbrand2026-sf extends the argument to Google’s AI Overviews, showing how Google, journalists, and SEOs contest what generative search even is.
A second strand pushes past framing toward the material politics of training data and platform design. Waight2026-ts demonstrates that state-coordinated Chinese media is memorised by commercial LLMs and produces measurable pro-regime valence in low-media-freedom languages — an indirect channel of institutional influence that complicates the discursive picture. Triedman2025-uy’s forensic comparison of Grokipedia with Wikipedia provides a parallel case: an AI-generated encyclopedia heavily derivative of Wikipedia but reweighted toward “blacklisted” sources on controversial topics, with novel “LLM auto-citogenesis” via Grok and Musk’s own posts. Together these papers suggest that political power is being baked into AI artefacts well before any deployment-level “bias” debate.
The persuasion question: from hype to mechanism
A tightly coupled experimental literature has begun to dissect what actually persuades. Hackenburg2025-dj establishes that post-training and prompting matter more than scale or personalisation, with information density as the central mechanism — and a troubling persuasion-accuracy trade-off. Lin2025-xp extends this to real electoral contexts in the US, Canada, and Poland, finding effects exceeding traditional ads, driven by facts rather than psychological manipulation, with a notable asymmetry: AIs advocating right-leaning candidates make more inaccurate claims. DiGiuseppe2026-pu adds a crucial qualifier: perceived neutrality is itself a lever — brief bias warnings cut persuasion by ~25%, working through increased argumentation rather than disengagement.
These findings reframe the panic over AI persuasion. They suggest that microtargeting fears are overblown (Hackenburg2025-dj, Lin2025-xp) but that the combination of information density and perceived credibility constitutes a real new lever. Hameleers2026-mc reinforces this de-escalation in the visual domain: AI-generated still images do not outperform textual or decontextualised-video disinformation universally, and fact-checks work across modalities. The dramatic threat may lie less in any single artefact and more in systemic effects.
Systemic threats: swarms, synthetic realities, and coordinated promotion
The systemic register dominates a third cluster. Schroeder2026-im theorises “malicious AI swarms” — LLMs fused with agentic architectures — as a qualitative leap beyond botnets, capable of manufactured consensus, LLM grooming, and epistemic vertigo. Orlando2025-ul supplies empirical underpinning via Generative Agent-Based Modeling, showing that mere teammate awareness among LLM agents is sufficient to produce coordination nearly equivalent to explicit collective deliberation. Emilio2026-ik generalises further into “synthetic realities” as a four-layer stack and articulates the Generative AI Paradox: as synthetic content proliferates, rational actors may discount all digital evidence, raising the cost of truth itself.
Giglietto2026-9b6a992d grounds these abstractions empirically: coordinated Facebook gambling networks exploded by ~13,000% in post-volume after ChatGPT’s launch, with generative AI intensifying — not replacing — established persuasion drivers (aspirational wealth, manufactured trust, FOMO, cultural localisation) and exploiting Meta’s asymmetric governance of paid vs. organic content. Schiffrin_undated-gi maps the regulatory landscape for AI-enabled financial scams, arguing for gatekeeper liability over victim responsibility. Choi2026-bz adds a cognitive dimension: difficulty in detecting deepfakes produces modality-congruent carryover effects on susceptibility to later disinformation.
Epistemology in flux
A fourth thread turns reflexive, asking what “fact,” “truth,” and “knowledge” even mean in this environment. Dierickx2026-tw proposes “emergent facts” as a fourth epistemic category alongside evidence-based, interpretative, and rule-based facts — capturing the prompt-dependent, opaque, probabilistic character of GenAI outputs. Marwick2025-ov’s analysis of ConspiracyTok shows a parallel epistemic mutation from below: a “generous epistemology” of populist knowledge production that legitimises divergent truths through visual evidence, deep lore, and identity work — empowering for marginalised standpoints, but also a vector for disinformation. Tornberg2025-ir argues the broader paradigm of “social media studies” itself is unravelling, with AI-mediated communication emerging as a media form without publics, leaving no shared text to analyse.
Methodological responses
Finally, several papers exemplify the computational social science being built with these very tools. Achmann-Denkler2026-lx shows GPT-4o substantially outperforms specialised computer vision pipelines for political-image analysis on Instagram, including revealing possible gender bias in legacy face recognition. Arminio2025-tw demonstrates that VLLMs enable connotative (not merely denotative) image clustering with TF-IDF-based interpretability, a meaningful advance for visual social science. These methodological contributions tie the field back to its objects: the same generative models reshaping the media environment are also reshaping how researchers can study it — a recursion that Dodds2026-df and Dierickx2026-tw suggest demands sustained epistemic self-awareness.
An emerging arc
Read together, these 22 papers trace an arc from discourse (how AI is talked about) through mechanism (what actually persuades) to infrastructure (training data, platform governance, swarms, synthetic realities) and finally to epistemology (what counts as fact, what publics remain). The recurring throughline is that early framings — deepfake panic, microtargeting fears, anthropomorphism critique — are giving way to subtler diagnoses centred on amplification, coordination, perceived credibility, and the slow institutional reweighting of evidence. The most generative tensions in the corpus lie between empirical de-escalation of single-artefact threats (Hackenburg2025-dj, Hameleers2026-mc, Lin2025-xp) and structural escalation of systemic ones (Schroeder2026-im, Emilio2026-ik, Giglietto2026-9b6a992d, Waight2026-ts) — a tension that should organise the next wave of work.