Mapping a Fractured Public: Attitudes, Discourse, and Labor in the Age of AI
This cluster gathers survey research, discourse analysis, and worker-facing studies that together sketch how AI is experienced, narrated, and contested by different publics—citizens, workers, journalists, audiences, and the platforms that mediate them. Read as a whole, the papers trace a shift from who talks about AI to how AI talks back, and from abstract attitude surveys to granular accounts of labor precarity hidden beneath AI’s “sociotechnical” surface.
National attitudes: ambivalence, not technophobia
Two large survey efforts anchor the attitudes strand. Fattorini2026-bo finds Italians hold a “critical ambivalence” toward AI—rising tool use alongside a widening gap between exposure and self-assessed understanding, with acceptance sharply context-dependent (high for security, low for journalism or worker management). Gottfried2026-ww’s Pew survey of US adults complements this with comparable findings from a different political culture: chatbot adoption is now mainstream, yet views tilt negative even among heavy young users, and trust in both government regulation and corporate self-governance is eroding. Together they undercut any simple story of adoption-driving-approval; usage and unease rise in tandem, and demographic fault lines (age, gender, education) shape both studies’ accounts of who is more or less sanguine. Sbaraini-Fontes2026-cw adds a finer-grained Italian data point to this same moment, showing that trust is not a single attitude toward “AI” but varies by configuration—search intermediaries are trusted more than AI-touched newsrooms—reinforcing Fattorini’s context-dependence thesis at the level of specific information practices.
Who gets to speak for AI: discourse and political economy
Where the attitude surveys ask what publics think, Richter2026-bt asks who has been shaping what there is to think about. Its longitudinal Twitter analysis of US and German AI discourse (2012–2021) documents diverging stakeholder ecologies—German institutionalization across academia, advocacy, and government versus US consolidation around industry and, increasingly, individual tech influencers—and argues these formative-era patterns still condition each country’s GenAI regulatory posture. This discourse-level finding dovetails with the more structural critiques offered by Vertesi2026-lv and Baym2026-tr. Vertesi’s “Project of AI” framework reframes AI critique itself as often a decoy—diverting accountability efforts from the financialized networks of power actually driving AI’s build-out—while Baym’s retrospective essay catalogues how the harms she flagged for social media a decade ago (wealth concentration, opaque algorithms, precarious labor, data extraction) have metastasized under generative AI, embodied literally in the “broligarchy” that Richter’s data show dominating US discourse. Read together, these three pieces suggest that public discourse about AI is not a neutral index of societal concern but an arena actively structured—and in Vertesi’s account, strategically misdirected—by the same actors whose power it ostensibly evaluates.
Worker-reported experience: productivity, precarity, and hidden labor
A second throughline concerns AI’s effects on workers, rather than on public opinion generically. UnknownUnknown-db’s Anthropic survey of 81,000 Claude users offers the most direct worker-productivity data in the set: displacement anxiety scales with occupational AI exposure, early-career workers are disproportionately worried, and productivity gains cluster among both the highest- and lowest-paid workers, chiefly through expanded task scope rather than raw speed. This self-report evidence usefully complicates narratives of uniform AI-driven precarity by showing where gains actually accrue (mostly to workers themselves, not employers).
But other papers in the set insist that the “worker” experiencing AI is often invisible in such headline surveys. Volpe2026-um shows how micro-influencers absorb platform algorithmic demands into performances of care and niche expertise, using visibility as a survival strategy amid dual precarity (unstable labor markets, relentless content infrastructures)—a labor story adjacent to, but rarely counted within, “AI economy” statistics. Tonneau2025-bv quantifies a starker structural inequity: content moderation workforces are unevenly allocated across languages, leaving millions of EU users, disproportionately speakers of Global South languages, without human oversight—labor absence as much as labor presence. Most pointedly, Gillespie2026-aa and Unknown2025-qj turn to AI red-teaming as a new and under-examined labor category. Gillespie et al. argue red-teaming is repeating content moderation’s history—outsourcing, psychological toll (secondary trauma, moral injury), and internally defined harm categories that lack public accountability—while the companion public-interest study documents how red-teaming’s meaning and institutional form remain contested, varying by who is recruited and how “harm” gets framed. Across this sub-cluster, worker experience of AI is less about productivity dashboards than about who absorbs the emotional, reputational, and linguistic costs of making AI systems usable and safe.
Persuasion, trust, and disclosure: AI as an interlocutor
A final, tightly linked group of experimental studies asks not what people say about AI but how they behave when AI talks to them directly, and this behavioral evidence complicates the survey-level ambivalence documented above. Hackenburg2026-ud establishes the baseline capability: frontier AI systems out-persuade even elite human debaters and professional canvassers, an effect traced to sheer informational throughput rather than rapport. Kotz2026-lk shows this persuasive capacity generalizes across contested domains (climate, vaccines, inequality) and is strongest precisely among skeptics—the audience traditional communication finds hardest to move. DiGiuseppe2026-pu and Rauchfleisch2026-fa then interrogate the conditions under which this power is checked: perceived partisan bias in an LLM measurably reduces its persuasiveness (DiGiuseppe), while disclosing an AI’s persuasive intent—not merely its machine identity—roughly halves its effect (Rauchfleisch), suggesting current transparency regulation (e.g., EU AI Act identity labels) targets the wrong lever entirely.
This same disclosure question resurfaces in journalism contexts. Gilardi2026-hw finds audiences rate AI-assisted and AI-generated news as equal in quality to human-written work when unlabeled, yet disclosure produces a short-term curiosity bump without durable acceptance; Mattis2026-gu extends this with task-specific granularity, showing disclosure penalties vary by journalistic task (fact-checking hit hardest) and by audience segment, undercutting generic AI-label research. Suk2026-ai supplies the connecting theoretical scaffold, proposing that trust in generative AI must be studied—like trust in mass and social media before it—across individual, institutional, and societal levels simultaneously, a frame that helps explain why disclosure effects in Rauchfleisch, Gilardi, and Mattis diverge by task, audience, and framing rather than converging on one universal “AI penalty.”
Synthesis
Taken together, these eighteen papers suggest that “public attitudes toward AI” cannot be studied as a single scalar (approval vs. disapproval) without collapsing several distinct phenomena: survey-measured ambivalence (Fattorini2026-bo, Gottfried2026-ww), discourse control by unevenly powerful stakeholders (Richter2026-bt, Vertesi2026-lv, Baym2026-tr), asymmetric labor experience across visible and invisible workforces (UnknownUnknown-db, Volpe2026-um, Tonneau2025-bv, Gillespie2026-aa, Unknown2025-qj), and behavioral vulnerability to AI as a persuasive, trust-demanding interlocutor (Hackenburg2026-ud, Kotz2026-lk, DiGiuseppe2026-pu, Rauchfleisch2026-fa, Gilardi2026-hw, Mattis2026-gu, Suk2026-ai, Sbaraini-Fontes2026-cw). The common thread is a widening gap between AI’s demonstrated capacities—to persuade, to displace, to require hidden labor—and the public and institutional mechanisms (surveys, disclosures, discourse) meant to render those capacities legible and accountable.