THE LONG VIEW

Bottom Line Up Front: Public AI information systems using open web search pose a significant risk to information integrity due to frequent citation of untrustworthy sources, despite improved answer coverage—posing a critical threat to informed democratic decision-making ahead of Iceland’s 29 August 2026 EU referendum [1]. Threat Identification: The integration of open web search in AI-powered public information services (e.g., Evrópuvefur) introduces unreliable or irrelevant sources into official responses, while curated retrieval systems, though more trustworthy, suffer from limited coverage and response refusal when knowledge gaps exist [1]. Notably, the system systematically omitted citations from RÚV, Iceland’s most trusted news source, indicating algorithmic bias or retrieval failure [1]. Probability Assessment: The risk is imminent and highly probable, with evidence showing 35% (65 of 187) of web-search answers containing flagged sources—almost always due to untrustworthiness or irrelevance—compared to rare flags in curated retrieval, which were limited to outdated content [1]. Weak prompt-level interventions (e.g., trusted-domain lists) showed minimal impact, increasing citations to approved domains from only 12% to 21%, indicating structural limitations [1]. Impact Analysis: The consequences include erosion of public trust in government-funded AI services, potential misinformation ahead of a constitutional decision, and diminished civic engagement. The invisibility of source quality to end users further amplifies risk, as fluency and topical fit do not correlate with trustworthiness, making poor-quality citations difficult to detect [1]. Recommended Actions: 1) Implement hybrid retrieval with mandatory source validation for web results; 2) Integrate authoritative local sources (e.g., RÚV) into retrieval indexing; 3) Deploy real-time source provenance transparency (e.g., trust ratings); 4) Conduct public audits of AI response quality in high-stakes contexts; 5) Re-evaluate prompt engineering efficacy and explore fine-tuning or retrieval re-ranking models for trusted domains [1]. Confidence Matrix: - Threat Likelihood: High (supported by 551 expert evaluations across 449 responses) - Impact Severity: High (affects democratic process and institutional credibility) - Mitigation Feasibility: Medium (requires technical and policy coordination) - Evidence Strength: High (empirical, expert-validated study with clear methodology) [1] [1] Hafsteinn Einarsson, Hafsteinn Birgir Einarsson, Jón Gunnar Ólafsson et al., 'Curated retrieval versus open web search in public AI information services: a coverage-trust trade-off,' arXiv:2607.03281 [cs.CY], 9 July 2026.