Utility-first deployment as a public trust strategy — the case for sequencing AI's civic contribution before its economic disruption
This paper proposes a deployment philosophy and institutional strategy, not a technical architecture. Its central claim — that sequencing AI deployment around visible public benefit builds durable social trust — is a hypothesis grounded in historical technology adoption patterns and public sentiment data. It is not a proven causal mechanism. The Foundation treats it as a governing strategic assumption subject to empirical testing, not as established fact.
Public trust in advanced AI systems will not follow automatically from their capabilities. It will be won or lost through the lived experience of the people who encounter them first — and the institutional context in which that encounter occurs. This paper argues that the entities most capable of shaping that context are not the major AI laboratories, whose incentives are primarily commercial, but the governance, nonprofit, civic, and standards institutions that can define an alternative deployment logic.
The paper proposes utility-first deployment as a coherent strategy: prioritizing AI applications that visibly reduce human fragility — in households, elder care, benefits navigation, civic access, and emergency preparedness — before prioritizing applications that maximize institutional efficiency or labor substitution. The argument is both strategic (utility-first deployment builds the social trust that enables more ambitious AI applications later) and ethical (reducing human fragility is a worthy end in itself).
The paper also names the legitimate criticisms of this position directly: utility-first deployment may be a delay tactic that does not change the trajectory of AI development; nonprofit governance frameworks lack enforcement authority; and the approach carries a paternalistic risk of deciding on behalf of others what beneficial AI looks like. These criticisms are addressed rather than dismissed. The paper's contribution is not to resolve them but to propose a governance architecture robust enough to remain credible despite them.
The term "AGI" is used throughout this paper and requires explicit definition because its meaning is contested and its conflation with current generative AI systems produces confusion about what is being governed.
For the purposes of this paper: current AI systems refers to large language models, multimodal models, and AI-assisted automation systems that exist and are deployed today. Advanced AI systems refers to systems significantly more capable than current systems, whether or not they satisfy any particular philosophical definition of general intelligence. AGI is used only when the argument specifically concerns systems whose capabilities are qualitatively different from current systems in ways that raise novel governance questions.
This distinction matters because the paper's governance proposals apply primarily to current and near-term AI systems, not to hypothetical AGI. The governance infrastructure the Foundation proposes to build — certification frameworks, deployment standards, verification architectures — is needed now, for systems that exist now. The AGI framing is relevant to the long-term strategic question of how public trust established now will or will not transfer to more capable future systems. It should not be used to imply that the paper's proposals are speculative or premature.
Similarly, utility-first deployment requires definition. The paper uses it to mean: prioritizing the sequencing of AI deployment in public-facing contexts toward applications where the primary measurable benefit is reduction of a specific human vulnerability or burden — accessibility barriers, information asymmetries, emergency preparedness gaps, or institutional complexity that currently disadvantages people without professional resources. It is not a claim that commercial AI applications are wrong or should be prohibited. It is a claim about sequencing and emphasis in the deployment contexts where governance institutions have influence.
Public concern about AI is documented and significant. Pew Research Center's June 2025 survey found that among U.S. adults who had heard of ChatGPT, 52% said the development of AI was more concerning than exciting, with only 17% more excited than concerned.1 Reuters/Ipsos polling in 2025 found that 71% of Americans worried AI could cause permanent job loss, and 61% believed AI poses risks to humanity's future.2
These figures describe concern about AI that already exists, not about hypothetical AGI. The concern is not primarily philosophical — it is experiential and economic. People are worried about job displacement, institutional surveillance, misinformation, and loss of human relevance in systems that are already affecting their lives.
The paper's central claim — that utility-first deployment changes these attitudes — is a hypothesis, not a demonstrated mechanism. The closest empirical analogies are instructive but not conclusive. Mobile banking adoption in Sub-Saharan Africa demonstrated that financial technology that directly reduced a specific household burden (access to financial services) achieved rapid adoption and high trust in populations that had been skeptical of financial institutions.3 Telemedicine adoption during COVID-19 demonstrated that medical technology framed as accessibility expansion rather than cost-cutting achieved faster trust development than prior telemedicine deployments framed around institutional efficiency.4 Neither analogy maps perfectly to AI governance, but both suggest that the framing and sequencing of benefit delivery matters to adoption and trust, not only the underlying capability.
The argument is not that beneficial deployment will make advanced AI safe. It is that visible, bounded, human-directed benefit creates the political and cultural conditions under which safety governance becomes possible — by giving people a reason to want AI to be governed rather than banned.
The paper also acknowledges what the evidence does not support: the claim that resistance to technology is simply a function of whether households benefit before institutions. The ATM eliminated bank teller positions with relatively modest public resistance.5 Electricity automation in manufacturing was bitterly resisted in some contexts and rapidly adopted in others depending on labor relations, regulatory context, and available alternatives. Technological transition resistance is more complex than any single frame captures. The utility-first argument is a hypothesis about one important variable, not a complete theory of technology adoption.
The commercial incentive structure of AI development is not designed to produce governance infrastructure as a primary output. The companies most capable of deploying advanced AI are optimizing for market capture, enterprise productivity, subscription revenue, and investor returns. These incentives are not uniformly hostile to beneficial deployment — many AI applications that are commercially successful are also genuinely useful — but they are not aligned with the specific governance problem this paper addresses.
The governance problem is: who sets the deployment standards that prevent AI systems from being used in ways that are technically legal but socially corrosive? Who certifies that a system claiming to provide unbiased public benefits information is actually unbiased? Who maintains the dispute infrastructure when an AI system's answer harms someone who relied on it? Who develops the professional standards for AI deployment in elder care, child services, and emergency response that protect vulnerable populations who lack the technical literacy to protect themselves?
These functions cannot be performed by the AI laboratories that develop the systems being governed, by the commercial entities that deploy them for profit, or by government agencies that lack the technical capacity and mandate to act proactively. They require independent institutions with a mandate for public benefit, technical credibility, and governance authority derived from demonstrated expertise rather than commercial scale or regulatory power.
A nonprofit foundation with genuine technical seriousness occupies a specific and valuable position in this ecosystem — not as a regulator and not as a developer, but as a standards and governance institution that can define what "aligned deployment" looks like concretely enough to be auditable.
The original paper's seven principles are reproduced here with additions: each principle is accompanied by implementation criteria — the observable conditions that would allow an independent auditor to determine whether a deployed system satisfies it.
The first public-facing applications of advanced AI in a given domain should be designed to reduce specific human burdens rather than eliminate human roles. Where AI can reduce the time, cost, or complexity of a task that humans currently perform with difficulty, prioritize that application before applications that substitute for human judgment entirely.
Implementation criteria: A compliant system can demonstrate a specific measurable burden reduced (time to navigate a benefits application, accuracy of form completion, accessibility for users with disabilities) that does not require eliminating a human decision point in a high-stakes domain.AI systems should not gain access to financial, health, legal, environmental, or personal data by default. Access requires explicit, informed, granular, revocable human authorization.
Implementation criteria: A compliant system can demonstrate an explicit permission architecture in which each data category is separately authorized, the authorization is logged, and revocation is technically possible and documented.Users should have access to what an AI system knows about them, what it is doing, why it is doing it, and what it is recommending — before they are asked to trust it with consequential decisions.
Implementation criteria: A compliant system provides a human-readable explanation of any significant output or recommendation, identifies the data sources used, and surfaces its confidence level and known limitations.When technically feasible, sensitive computations and sensitive data should be processed locally or under user-controlled infrastructure rather than transmitted to centralized commercial servers.
Implementation criteria: A compliant system documents which computations are performed locally, which require cloud transmission, what data is retained after a session, and the user's options for limiting cloud transmission.AI-generated outputs in domains where errors have significant consequences for health, legal status, financial security, or safety must carry explicit review status — and must not be presented with the authority of reviewed outputs until they have been reviewed.
Implementation criteria: A compliant system applies the ARIA Network verification status taxonomy or an equivalent published standard; outputs without expert review are labeled as such; high-stakes domains require human review before outputs carry authority status.AI deployment decisions should account for the fragility introduced alongside the efficiency gained. A system that optimizes household energy efficiency while creating a single point of failure for heating in cold weather has traded resilience for efficiency in a way that may not be in the household's interest.
Implementation criteria: A compliant system documents failure modes, maintains manual override capability for all safety-relevant functions, and does not eliminate non-AI backups as a condition of adoption.AI systems deployed in contexts involving significant human relationships — elder care, child services, mental health support, grief navigation, disability assistance — should preserve the human relationship as the primary site of meaning, with AI functioning as a support layer rather than a substitute for human presence and judgment.
Implementation criteria: A compliant system in a human-care context cannot be deployed as the primary relationship interface for vulnerable users without documented human oversight, escalation pathways to human professionals, and explicit design decisions against dependency optimization.Figure 1 — The relationship between the three ARIA papers. Transitional AGI Governance (top) is the governing philosophy — it defines what aligned deployment means and which principles applied frameworks must satisfy. ARIA Home (left) applies those principles to residential AI environments. ARIA Network (right) applies them to civic knowledge environments. Both share a certification infrastructure (bottom) governed by the same anti-capture architecture.
ARIA stands for: Adaptive (responsive to user needs and context), Responsible (governed by the seven principles above), Interoperable (open standards, not vendor lock-in), Accountable (logged, auditable, and correctable). These are not branding choices — they are architectural requirements that the Foundation's applied proposals must satisfy to carry the ARIA designation.
ARIA Home demonstrates principles 1 (help before replacing — safety and accessibility without eliminating human control), 2 (permission before access — the seven-class permission taxonomy), 3 (visibility before trust — the audit log), 4 (local before extractive — local-first processing), and 6 (resilience before efficiency — manual override requirements) applied in the residential context.
ARIA Network demonstrates principles 3 (visibility before trust — verification status labels), 5 (review before authority — the five-tier verification system), and 7 (human dignity before automation — expert review requirements in high-stakes domains) applied in the civic knowledge context.
A nonprofit cannot sustain governance work on moral seriousness alone. The Foundation's proposed revenue streams carry their own risks, which must be named rather than obscured by optimistic projections.
| Revenue Stream | Estimated Year 3 Potential | Primary Risk | Mitigation |
|---|---|---|---|
| ARIA-Ready device certification | $250K–$500K (50–100 manufacturers at $2.5K–$5K) | Capture — manufacturers who fund certification influence standards | 15% revenue concentration limit; standards committee independent of fee revenue; public disclosure of all manufacturer relationships |
| ARIA Network agent certification | $100K–$200K (100–200 agent operators at $500–$2K) | Gaming — operators optimize for certification criteria rather than genuine alignment | Annual re-certification with updated criteria; spot-check audits; public accuracy reporting requirements |
| Installer and integrator training | $50K–$150K (training program licensing) | Quality dilution — training certifications that don't reflect actual competence | Practical competency assessment; failure rate reporting; third-party examiner option |
| Institutional consulting | $100K–$300K (eldercare, housing, civic org deployments) | Mission drift — consulting priorities displacing research priorities | Board-approved cap on consulting as percentage of revenue; consulting must align with active Foundation research areas |
| Grants — public interest | $200K–$500K (federal, foundation, civic grants) | Agenda distortion — grant priorities may not align with Foundation's assessment of most important work | Foundation accepts grants only for work already on the research agenda; no work-for-hire grant structures |
| Expert review marketplace fees | $50K–$150K (percentage of expert reviewer payments) | Exploitation of reviewers; credential inflation pressure | Published reviewer compensation minimums; credential standards not subject to platform commercial pressure |
| API access — verified knowledge records | $50K–$100K (institutional API subscribers) | Data aggregation privacy risks; competitive pressure to expand data scope | Data minimization by design; GDPR-compliant erasure; Foundation board approval for any new data category |
| Estimated Year 3 Total Range | $800K–$1.9M — sufficient for 8–15 staff and ongoing research program | ||
| Phase | Timeline | Key Deliverables | Success Criteria |
|---|---|---|---|
| Phase 1 Concept Publication | Complete — May 2026 | Three foundational papers (ARIA Home, ARIA Network, Transitional AGI Governance) published with open licensing | Papers cited or engaged by at least three external researchers, policymakers, or institutions within 6 months |
| Phase 2 Prototype and Demonstration | Q3–Q4 2026 | ARIA Home reference prototype; ARIA Network closed pilot (three boards); agent registry template; verification taxonomy deployed | ARIA Home prototype demonstrated to at least one institutional partner (eldercare, housing, disability org); ARIA Network pilot completes 3-month run with published findings |
| Phase 3 Certification Launch | Q1–Q2 2027 | ARIA-Ready certification program launched with published criteria; first batch of certified devices or systems; Foundation publishes first annual transparency report | Minimum 10 certified systems; at least one institutional deployment referencing ARIA-Ready in procurement criteria; first annual report published on schedule |
| Phase 4 Partnerships | 2027 | Formal partnerships with universities, legal aid organizations, eldercare organizations, and at least one civic technology group | At least one joint research publication; at least one partner deploying ARIA-certified infrastructure; at least one academic institution incorporating Foundation standards in coursework |
| Phase 5 Revenue and Scaling | 2027–2028 | Certification revenue funding ongoing operations; expert verification marketplace operational; annual ARIA governance report | Revenue covers 60% of operating budget without single-source concentration above 15%; Foundation publishes audited financials; governance board established with independent members |
The central empirical claim is unproven. The paper argues that utility-first deployment builds social trust that enables more ambitious AI governance later. This is a plausible hypothesis grounded in historical analogies, not a demonstrated causal mechanism. The Foundation should treat it as a testable assumption and design its program evaluation accordingly.
The scope is necessarily limited. Everything the Foundation does in the 5-year implementation roadmap above will affect a small fraction of global AI deployment. This is not an argument against doing it — standards institutions begin with limited scope and grow through demonstrated value — but it should not be obscured by language implying the Foundation is governing the AI transition at civilizational scale.
The seven principles require governance processes to develop and maintain. Principles without ongoing governance become dated. The Foundation needs a formal review process that updates implementation criteria as AI capabilities change — otherwise the standard becomes a compliance exercise rather than a living governance instrument.
Without utility-first framing and governance infrastructure developed by independent institutions, the default narrative of the AI transition continues to be shaped primarily by the commercial interests of major AI laboratories and the political interests of governments focused on national competitiveness. The households, elder care residents, disabled users, and civic communities most dependent on AI governance are the least represented in that narrative. The Foundation's contribution is not to resolve this problem — it is to insert a coherent alternative institutional voice into the conversation while there is still a conversation to participate in.
How should the seven principles be weighted when they conflict — when maximum accessibility requires cloud processing that sacrifices local-first privacy? What is the appropriate mechanism for including affected communities in standards development, and how are those processes governed? How does the Foundation maintain the distinction between governance institution and advocacy organization as it grows? What is the success condition at which the Foundation should consider itself to have achieved its transitional governance mission, and what happens after that?
The Foundation's governance implications for the broader AI policy ecosystem: establish that independent certification is achievable and credible; demonstrate that nonprofit institutions can develop technically serious AI governance; create procurement and insurance reference standards that make ARIA compliance commercially advantageous; and publish all governance architecture without restriction so that the standard's value is not dependent on Foundation survival.
If Foundation-certified AI deployments demonstrate no measurable difference in public trust outcomes compared to non-certified deployments — measured by longitudinal survey at 24 months — the utility-first hypothesis is not producing its claimed benefit and the certification program requires fundamental redesign.
If the revenue model proves unable to sustain operations without exceeding the 15% single-source concentration limit, the anti-capture architecture will require revision — either loosening the concentration limit with compensating governance safeguards, or accepting that the Foundation's scope must be reduced to match its independent revenue capacity.
If major AI deployers systematically adopt the utility-first framing without adopting the governance infrastructure — using the language of household benefit and human dignity as marketing while deploying without audit logs, permission architecture, or verification standards — the framing strategy will have produced cultural cover for misaligned deployment rather than governance compliance.
Final Clarification
The governance gap in the AI transition is not primarily technical. The technical problems are being worked on by well-resourced laboratories. The governance gap is institutional: there is no constituency of independent institutions with the technical credibility, public mandate, and governance authority to define what aligned deployment looks like in the domains where most people will first encounter advanced AI.
The Foundation's ambition is not to govern all of AI. It is to occupy that gap — to be one institution among several that can say, with demonstrated credibility: this is what alignment looks like in practice, and here is how to audit it.
Advanced intelligence must first prove that it can help humanity become less fragile. The Foundation exists to define what that proof looks like.
One-Page Policy Brief
For policymakers, funders, and institutional partners — EM Foundation, May 2026
The Problem
52% of US adults who have used AI tools say development is more concerning than exciting (Pew, 2025). 71% worry about permanent job loss (Reuters/Ipsos, 2025). This is not irrational fear — it reflects a deployment pattern where institutions capture efficiency gains before households feel protected.
The Opportunity
Mobile banking in Sub-Saharan Africa and telemedicine adoption during COVID-19 both demonstrated that technology deployment framed around reducing specific household burdens achieves faster trust development and adoption than deployment framed around institutional efficiency.
The Proposal
Governance institutions should prioritize AI deployment standards that sequence public-facing applications around measurable burden reduction — elder care, benefits navigation, disability access, civic information — before optimizing for labor substitution or institutional efficiency. This builds the political and cultural conditions for AI governance to succeed.
For Policymakers
Require AI procurement criteria to include ARIA-Ready or equivalent certification for government-funded AI in public-benefit services. Create liability frameworks that make audit logs and permission architecture commercially advantageous.
For Funders
Fund governance infrastructure development alongside AI capability development. The ratio of public investment in AI capabilities versus AI governance is currently severely imbalanced. Governance institutions need operating capital, not only project grants.
For Partners
Eldercare organizations, disability advocacy groups, legal aid institutions, and civic technology organizations that adopt ARIA-certified deployments demonstrate the standard in practice and create the deployment precedents that governance frameworks need to become credible.
Full paper: emfoundation.net/paper-transitional-agi-governance.html · Contact: research@emfoundation.net · Published without copyright restriction