National Publics and the Texture of AI Attitudes

The baseline empirical picture of public opinion comes from large-scale surveys. Gottfried2026-ww’s Pew study of US adults and Fattorini2026-bo’s Italian Observa Monitor data converge on a similar diagnosis despite different national contexts: adoption of AI tools (chatbots, smart devices, search summaries) is rising rapidly, yet attitudes are simultaneously hardening into skepticism or what Fattorini calls “critical ambivalence” rather than technophobia. Both studies show this skepticism is not simply a function of unfamiliarity—heavy users (younger Americans, Italian chatbot adopters) are often among the most pessimistic about societal-level impact even as they report personal convenience. Both also find durable demographic cleavages: gender gaps in risk perception, age-based divergence in trust toward institutions governing AI, and acceptance that is highly domain-dependent (high for security/efficiency uses, low for journalism, hiring, and creative work). Suk2026-ai offers a conceptual scaffolding for this body of survey work, arguing that trust in generative AI should be theorized as a multilevel phenomenon—individual, institutional, societal—borrowing from decades of mass-media and social-media trust research rather than treated as sui generis. Read together, these three papers suggest the public is not undergoing a binary embrace-or-reject process but is actively renegotiating, issue by issue, what kind of intermediary AI is allowed to be.

Trust, Disclosure, and the Information Ecosystem

A cluster of experimental studies narrows this general trust question to the specific arena of AI-mediated information and news. Gilardi2026-hw and Mattis2026-gu both use disclosure experiments but reach subtly different emphases: Gilardi finds that quality perceptions of AI-written news are statistically indistinguishable from human-written news until disclosure occurs, at which point a curiosity-driven bump in short-term engagement appears without translating into durable acceptance; Mattis, testing task-specific disclosures along the journalistic value chain, finds disclosure effects are uniformly negative but vary sharply by task (fact-checking penalized most, idea/image generation least) and by individual moderators such as political position and AI literacy—undercutting the idea of a single, generic “AI label” effect. Sbaraini-Fontes2026-cw extends this logic comparatively, showing Italian audiences trust AI-generated search results more than AI-involved news, implying that where AI sits in the information chain—intermediary versus producer—shapes legitimacy independently of the content itself. Dubey2026-bl’s chatbot studies complicate assumptions in this literature from another angle: rather than distrust tracking conspiracy belief as expected, high-conspiracy users actually trusted and intended to use “balanced news” chatbots more than low-conspiracy users, suggesting perceived neutrality and usefulness can override source-level suspicion in ways current disclosure-effect models do not anticipate.

Persuasion and the Politics of AI Credibility

Two papers push from attitude-toward-AI into AI’s attitude-changing capacity. Hackenburg2026-ud demonstrates that frontier LLMs reliably out-persuade even elite, incentivized human persuaders—debate champions, professional canvassers—and that this advantage is driven by information throughput (fact density) rather than rapport, generalizing from survey attitudes to real monetary donations. This raises the stakes considerably: AI is not just an object of public evaluation but an increasingly potent persuasive agent in its own right. DiGiuseppe2026-pu supplies a crucial qualifier: this persuasive power is politically contingent. A brief cue that the model is “biased” against one’s party reduces belief correction by roughly a quarter, operating via increased counter-argument rather than disengagement—evidence of motivated reasoning transferring directly from human-source contexts to AI sources. Together these two studies suggest an emerging two-level finding: AI’s raw persuasive capacity may exceed that of trained humans, but its real-world efficacy remains hostage to perceived neutrality, meaning public trust (studied in the cluster above) is not a separate concern from AI’s persuasive power but a direct moderator of it.

Worker-Reported Productivity, Displacement, and Platformized Labor

A parallel strand shifts from citizens-as-audiences to workers-as-users. The Anthropic report UnknownUnknown-db, based on ~81,000 Claude.ai users, offers the topic’s most direct self-report data on productivity and displacement: exposure to AI correlates with displacement anxiety, productivity gains concentrate among the highest- and lowest-paid occupations, and gains typically accrue to workers themselves rather than employers—except notably for early-career workers, who report both less personal benefit and more displacement concern. This early-career vulnerability echoes a smaller-scale but conceptually related finding in Cotter2019-js’s study (an outlier in method and era) of Brazilian tattoo artists on Instagram, where platform dependence reshapes workspace, procedure, and unpaid labor time regardless of AI per se—illustrating a longer lineage of “platformized work” that generative-AI labor studies are now inheriting and intensifying. Where the Anthropic report asks what workers gain, Gillespie2026-aa and Tonneau2025-bv ask what certain workers absorb on the system’s behalf: red-teamers face secondary trauma and moral injury performing adversarial, often invisible safety labor, while content moderators are unevenly allocated across languages, leaving many Global South and non-English-speaking users essentially unmoderated. Moran2025-qn rounds out this arc by documenting the professional precarity of the trust-and-safety field itself—the institutional apparatus meant to manage these harms is being hollowed out by layoffs and political pressure just as the demands on it (per Gillespie and Tonneau) intensify. Read in sequence, these four papers trace a single argument: AI’s labor costs are not evenly distributed productivity gains but a stratified system in which some workers reap scope-expansion benefits while others—moderators, red-teamers, safety professionals, platform-dependent gig workers—absorb the invisible costs of keeping the system usable.

Critical Counter-Currents: Political Economy and Reflexivity

Two essays step back from survey and experimental findings to interrogate the frame itself. Vertesi2026-lv argues that much accountability research—including, implicitly, work on disclosure, trust, and red-teaming surveyed above—risks functioning as a “decoy” that individualizes attention onto algorithmic particulars while the real action is financialization and network-power consolidation by AI firms; even worker-harm narratives (the “disruption decoy”) can obscure corporate restructuring and offshoring rather than illuminate it. Baym2026-tr makes a kindred but more autobiographical case: a decade of critical scholarship on platforms’ wealth concentration, algorithmic opacity, precarious labor, and data extraction has coincided with all four conditions worsening, generative AI now compounding extraction (training on user content, producing deepfakes) atop the earlier social-media harms. Both pieces function as a methodological conscience for the rest of the topic: they do not dispute the empirical findings on attitudes, trust, or displacement reported elsewhere in this note, but insist that documenting public perception and worker experience is necessary yet insufficient—that without attention to who owns the infrastructure and captures the value, survey-based diagnosis risks becoming, in Vertesi’s term, co-constitutive of the very Project it studies rather than a check on it.