Public Attitudes and Labor Experiences with AI
National surveys and the texture of ambivalence
The most direct evidence on public sentiment comes from large-scale, nationally representative surveys. Gottfried2026-ww’s Pew study of US adults and Fattorini2026-bo’s analysis of Italy’s Observa Monitor both trace a similar arc: rapid growth in everyday AI use (chatbots, smart devices, search summaries) running alongside flat or worsening societal outlooks. Pew documents that roughly half of US adults now use AI chatbots, yet views tilt negative even among young, heavy-using cohorts; Fattorini finds that Italians’ sense of AI as a “threat to humanity” rose sharply between 2023 and 2024 even as tool adoption climbed. Both studies converge on a picture of use without conviction—adoption outpacing trust—and both surface persistent demographic fault lines, with women and older or less-educated respondents expressing more caution in both countries. Fattorini’s framing of “critical ambivalence” (support for AI in security contexts, rejection in journalism, hiring, and art) gives conceptual shape to what Gottfried’s numbers show empirically: acceptance is domain-contingent, not a stable disposition toward “AI” as a monolith. Suk2026-ai offers a theoretical bridge for this terrain, proposing that trust in generative AI be studied as a multilevel phenomenon—individual, institutional, societal—continuous with, but not reducible to, decades of media-trust research; it functions as connective tissue between the country-level surveys and the more granular journalism-trust experiments below.
Trust, disclosure, and the mechanics of AI-mediated information
A cluster of experimental papers zooms into one specific site of public attitude formation: news and information produced or mediated by AI. Gilardi2026-hw finds that Swiss readers rate AI-written and human-written news as equivalent in quality when authorship is undisclosed, and that disclosure produces only a short-lived curiosity bump rather than durable acceptance. Mattis2026-gu complicates this with a more granular, task-specific lens: every disclosed use of generative AI in the Dutch news value chain reduces trust, but unevenly—idea generation is forgiven, fact-checking is punished—and the effect is moderated by political position and audience knowledge rather than topic. Sbaraini-Fontes2026-cw extends the comparison across contexts, showing Italian audiences trust AI-generated search results more than AI-involved news, suggesting that the configuration in which AI appears (intermediary versus institutional author) shapes trust independent of the underlying technology. Together these three papers argue against a single “AI aversion” story: trust responses are structured by task, context, and disclosure design, echoing Suk’s call for multilevel theorizing and giving it empirical flesh.
Persuasion, transparency, and the politics of perceived neutrality
A second experimental thread asks not how audiences judge AI-produced content but how AI acts on their beliefs. Hackenburg2026-ud establishes the baseline capability: frontier conversational AI out-persuades even elite, coached human debaters and professional canvassers, an advantage traced to sheer information throughput (fact density) rather than rapport, and one that translates into real monetary donations. Rauchfleisch2026-fa and DiGiuseppe2026-pu then interrogate the conditions under which this persuasive power can be blunted. Rauchfleisch shows that EU-AI-Act-style identity labels do almost nothing to curb persuasion, but disclosing the chatbot’s persuasive intent roughly halves it—relocating the locus of meaningful transparency from “what the system is” to “what it is trying to do.” DiGiuseppe approaches the same fragility from the demand side: telling users an LLM is politically biased against their party reduces belief correction by up to 28%, not because users disengage but because they argue back more, a motivated-reasoning response rather than simple discounting. Read together, these three papers sketch a coherent finding: AI persuasion is real and large, but it is not brute or context-free—it is contingent on perceived source neutrality and disclosed intent, giving policymakers actual levers (short of banning AI persuasion outright) for governance.
Worker-side accounts: productivity, displacement, and the hidden labor beneath the interface
Where the preceding sections examine AI from the audience/citizen side, UnknownUnknown-db and Gillespie2026-aa turn to workers—those using AI and those constructing it. Anthropic’s survey of ~81,000 Claude users, in UnknownUnknown-db, finds that self-reported productivity gains and displacement anxiety are not opposites but travel together: workers in highly AI-exposed occupations report both the largest productivity gains and the greatest fear of displacement, with a U-shaped relationship between speedup and perceived job threat, and with gains accruing disproportionately to workers themselves rather than employers. This complicates the anxiety/optimism narratives in the general-public surveys (Gottfried, Fattorini) by grounding them in occupation-specific, task-level self-report. Gillespie2026-aa pushes the labor lens further upstream, into the production of AI safety itself: red-teaming, the argument goes, is a sociotechnical labor system—echoing content moderation’s history of outsourcing, opaque value-setting, and psychological harm to workers—that is currently invisible to the very publics whose attitudes the other papers measure. Read against UnknownUnknown-db, it suggests that “worker experience of AI” spans a spectrum from white-collar productivity users self-reporting speedups to precarious ghost workers whose adversarial labor makes the models trustworthy in the first place, and that public-attitude surveys largely miss this latter population entirely.
The political economy frame: what these attitudes and experiences might be occluding
Vertesi2026-lv provides a critical counter-frame to the entire topic cluster. It argues that survey-measured attitudes, disclosure experiments, and even labor-focused critiques of displacement or red-teaming risk becoming “decoys”—debates over safety, disruption, or regulatory design that inadvertently legitimize a deeper “Project of AI” organized around capital accumulation and network power. This is a productive friction with the rest of the topic: the fine-grained findings on trust, persuasion, and displacement documented elsewhere (e.g., in Gottfried2026-ww, UnknownUnknown-db, Rauchfleisch2026-fa) are exactly the sort of “disruption” and “safety” framings Vertesi warns can distract from questions of who owns the infrastructure and who captures the value. Placed at the end of this arc, Vertesi’s essay asks the other eleven papers a pointed question: are we measuring public attitudes and worker experience as ends in themselves, or as symptoms of a political-economic realignment that the survey and disclosure literatures are not designed to see?