Abstract
The transition to superintelligence will outpace today’s policy toolkit, so democratic societies need an ambitious new industrial policy — comparable to the Progressive Era and New Deal responses to industrialisation — that keeps the AI economy open and participatory while building institutions capable of managing frontier risks.
Framing
- AI has progressed from narrow, minutes-long tasks to general tasks taking hours; if progress continues, systems will handle projects that currently take people months.
- Superintelligence is defined as AI capable of outperforming the smartest humans even when those humans are assisted by AI.
- Three stated goals for the transition:
- Share prosperity broadly — rising living standards, lower costs, wider access to AI-driven opportunity rather than benefits captured by a few.
- Mitigate risks — economic disruption, misuse in cyber and biology, loss of alignment or control; safety must scale with capability.
- Democratize access and agency — participation in the AI economy should not require access to the most powerful models, only to AI that is useful, affordable and privacy-preserving.
- Risks named explicitly: job and industry disruption, misuse by bad actors, misaligned systems evading control, governments deploying AI against democratic values, and concentration of power and wealth.
The case for industrial policy
- Markets are treated as the default; industrial policy is justified where new technologies create opportunities and risks that existing institutions cannot manage.
- Historical precedent: the Progressive Era and New Deal modernised the social contract for electricity, the combustion engine and mass production via labour protections, safety standards, safety nets and expanded education.
- Proposed toolbox: research funding, workforce development, market-shaping tools and targeted regulation.
- Division of labour: nongovernmental institutions pilot, measure and iterate; governments scale what works through procurement, regulation and investment — framed as a hedge against regulatory capture and centralised control.
- Near-term steps flagged as already necessary: AI data centres should pay their own way on energy so households do not subsidise them, and should generate local jobs and tax revenue; common-sense regulation should protect children and address national security risks.
- The document is scoped to the United States as a starting point, and is presented as early and exploratory rather than a final set of recommendations.
Section 1 — Building an open economy
Voice and entrepreneurship
- Worker perspectives — a formal mechanism for workers to collaborate with management on AI deployment, letting workers prioritise deployments that remove dangerous, repetitive, administrative or exhausting tasks, and setting limits on uses that intensify workloads, narrow autonomy or undermine fair scheduling and pay.
- AI-first entrepreneurs — use AI to absorb the overhead that blocks entrepreneurship (accounting, marketing, procurement), paired with microgrants or revenue-based financing and “startup-in-a-box” supports such as model contracts and shared back-office infrastructure. Worker organisations could provide training, shared services and help negotiating commercial terms and IP.
- Right to AI — treat AI access as foundational infrastructure, analogous to literacy campaigns or electricity and internet rollout. Includes affordable access to foundational models, a broadly available baseline capability with free or low-cost access points, plus the education, connectivity and training to use them.
Fiscal and ownership mechanisms
- Modernize the tax base — AI may expand corporate profits and capital gains while shrinking labour income and payroll taxes, eroding funding for Social Security, Medicaid, SNAP and housing assistance. Proposed rebalancing toward capital-based revenues (higher capital gains taxes at the top, corporate income, targeted measures on sustained AI-driven returns), plus exploration of taxes on automated labour, paired with wage-linked incentives modelled on R&D credits to encourage firms to retain and retrain workers.
- Public Wealth Fund — a fund giving every citizen a stake in AI-driven growth, invested in diversified long-term assets covering both AI companies and firms adopting AI, with returns distributed directly to citizens. Seeding to be determined jointly by policymakers and AI companies.
- Accelerate grid expansion — public-private partnership models to finance energy infrastructure for AI, targeting financing constraints, permitting delays and siting risk for interstate and interregional transmission. Tools include investment credits, flexible subsidies, equity stakes, removal of barriers to advanced conductors and HVDC, and a narrow federal authority to accelerate interregional transmission in the national interest. Structured to limit taxpayer exposure and to lower household energy costs.
Economic security
- Efficiency dividends — convert AI-driven cost savings into worker benefits: larger retirement matches, greater employer share of healthcare costs, subsidised child and eldercare. Also time-bound 32-hour/four-day workweek pilots with no loss of pay holding output constant, converting reclaimed hours into a permanent shorter week or bankable paid time off, plus “benefits bonuses” tied to measured productivity.
- Adaptive safety nets — first make unemployment insurance, SNAP, Social Security, Medicaid and Medicare fully functional and responsive; then invest in real-time measurement of AI’s effect on work, wages, job quality and sectoral dynamics; then define expanded temporary supports (flexible unemployment benefits, fast cash assistance, wage insurance, training vouchers) that trigger automatically when metrics cross pre-defined thresholds and phase out as conditions stabilise, avoiding permanent programme expansion.
- Portable benefits — healthcare, retirement savings and skills training in accounts attached to the individual rather than the job, pooling contributions from multiple sources; retirement modernised through pooled structures allowing continuous accrual across employers.
- Pathways into human-centered work — expand the care and connection economy (childcare, eldercare, education, healthcare, community services) as an absorber of displaced workers, supported by training pipelines and incentives for employers to raise pay and conditions. Complemented by a family benefit treating caregiving as economically valuable and compatible with part-time work, retraining or entrepreneurship.
Science infrastructure
- AI-enabled laboratories — a distributed network to test and validate AI-generated hypotheses at scale, integrating AI into experimental workflows through automation of routine processes, high-quality data capture and rapid hypothesis-test iteration.
- Paired with physical infrastructure to translate validated discoveries into deployment, aligned financing, and sustained investment in training scientists, technicians and operators.
- Explicit instruction to distribute both lab and production infrastructure across universities, community colleges, hospitals and regional research hubs rather than concentrating it in elite institutions.
Section 2 — Building a resilient society
Framing
- Risks named: misuse for cyber or biological harm, pressure on social and emotional wellbeing (particularly for young people), misalignment with human intent, operation beyond meaningful oversight, and strain on the institutions that keep societies stable and free.
- Existing upstream safeguards — global standards, transparency on evaluations and mitigations, model testing, red teaming, usage policies, the EU AI Act and US state regulation — should continue, but resilience increasingly depends on what happens after deployment.
- Analogies drawn to electrical safety standards, automobile safety systems, aviation monitoring and coordinated response, and post-market surveillance in food and medicine — with the caveat that those systems were built with the luxury of time.
Safety and verification
- Safety systems for emerging risks — tools to protect models, detect risks and prevent misuse in high-consequence domains; advanced AI applied to threat modelling, red teaming, net assessments and robustness testing; complementary protective systems such as rapid medical countermeasure production and expanded strategic stockpiles. Competitive safety markets to be catalysed through procurement, standards, insurance frameworks and advance-purchase commitments.
- AI trust stack — provenance and verification standards, secure verifiable signatures for actions such as generating content or issuing instructions, and privacy-preserving logging and audit systems that support investigation without enabling pervasive surveillance. Includes governance frameworks clarifying how accountability is assigned to roles and how delegation, monitoring and escalation function.
- Auditing regimes — strengthen bodies such as the Center for AI Standards and Innovation (CAISI) to develop frontier-risk auditing standards with national security agencies, and use procurement, advance-purchase commitments, insurance and standards-setting to build a competitive market of auditors. Standards designed for international adoption to reduce fragmentation.
- A narrow set of highly capable models — particularly those materially advancing chemical, biological, radiological, nuclear or cyber risk — may eventually require pre- and post-deployment audits, applied only to a small number of companies and the most advanced models.
- Model-containment playbooks — coordinated, tested procedures for containing dangerous systems already released, covering cases where weights are public, developers are unwilling or unable to restrict access, or systems are autonomous and self-replicating. Drawn from cybersecurity and public health experience that coordinated action reduces impact even without full containment.
Governance and accountability
- Mission-aligned corporate governance — frontier developers should adopt structures embedding public-interest accountability, such as Public Benefit Corporations, with explicit commitments to broadly shared benefits including long-term philanthropic giving. Also hardening against corporate or insider capture: securing model weights and training infrastructure, auditing models for manipulative behaviour or hidden loyalties, and monitoring high-risk deployments.
- Guardrails for government use — codified rules for government AI use with high reliability, alignment and safety standards, reinforced technically. Conversely, AI-assisted government workflows create clearer digital records that inspectors general, congressional committees and courts could audit with AI-enabled tools. Transparency frameworks including FOIA should be modernised, clarifying when AI-interaction and agentic action logs constitute retainable federal records.
- Mechanisms for public input — alignment should not be defined solely by engineers or executives; developers should publish model specifications and evaluation information, while governments anchor standards in democratic law and establish representative public input processes alongside business stakeholders.
Coordination
- Incident reporting — a mechanism for companies to report incidents, misuse and near-misses to a designated public authority, emphasising learning over punishment, with scoped public disclosure protecting technical, national security and competitive information. Near-misses would include concerning internal reasoning or unexpected capabilities even where safeguards held.
- International information-sharing — expand CAISI as a trusted technical evaluation body, then build a global network of AI Institutes sharing protocols, joint evaluations and coordinated mitigations, potentially evolving into a multilateral safety and standards institution with secure cross-lab and cross-country channels and crisis communication. Requires antitrust safe harbours and narrowly scoped information-sharing rules, and should cover societal risks such as youth safety, not only national security.
Next steps announced
- Feedback channel at newindustrialpolicy@openai.com.
- A pilot programme of fellowships and focused research grants of up to $100,000 and up to $1 million in API credits for work building on these ideas.
- Discussions convened at a new OpenAI Workshop opening in May in Washington, DC.