Industrial Policy for the Intelligence Age: Ideas to Keep People First
Published April 2026 (PDF cover); June 9, 2026 (landing page, marked as an update) · Two dates are printed, for two different things. The PDF's cover page reads "April 2026" and nothing inside the PDF carries a later date; its metadata gives a creation date of 3 April 2026. The openai.com landing page that links the PDF shows "June 9, 2026" in its dateline because it was edited that day to add an "Update from June 9" paragraph saying the feedback inbox had closed after more than 400 responses and grant recipients were under review. The document text itself was not revised, so the sort date is April 2026. The landing page also mentions the OpenAI Workshop "opening in May in Washington, DC", which only reads correctly from an April vantage point.
Not law. This is a company's own public position on AI regulation. It is not law, and it carries no legal force.
What it argues for
This is OpenAI's broadest statement of what governments should do about the economic and social consequences of advanced AI, rather than a narrow frontier-safety position. It frames the moment as "a transition toward superintelligence" and says its purpose is "to start a conversation about governing advanced AI in ways that keep people first", explicitly offering the ideas as "intentionally early and exploratory" and focused "on the United States as a starting point". Its near-term asks are concrete: "AI data centers should pay their own way on energy so that households aren’t subsidizing them", and "Governments should implement common-sense AI regulation—not to entrench incumbents through regulatory capture but to protect children, mitigate national security risks, and encourage innovation." Beyond that it argues for an industrial-policy agenda on the grounds that "In normal times, the case for letting markets work on their own is strong" but these are not normal times, and sets out two blocks of proposals. The "Open Economy" block asks for worker voice in AI deployment, a "Right to AI", rebalancing the tax base toward capital-based revenues and possibly "taxes related to automated labor", a Public Wealth Fund giving every citizen a stake in AI-driven growth, 32-hour-week pilots, safety nets that trigger automatically on displacement metrics, and portable benefits. The "Resilient Society" block asks for strengthening the Center for AI Standards and Innovation (CAISI) to develop frontier auditing standards, pre- and post-deployment audits applied "only to a small number of companies and the most advanced models", incident and near-miss reporting to a designated public authority that should "emphasize learning and prevention over punishment", model-containment playbooks, Public Benefit Corporation-style governance for frontier labs, statutory guardrails on government use of AI, and an international network of AI institutes with antitrust safe harbours for safety information-sharing. The organising principle is stated in one line: "As capability scales, safety must scale with it."
Stated positions (13)
- Near-term regulation should be "common-sense AI regulation—not to entrench incumbents through regulatory capture but to protect children, mitigate national security risks, and encourage innovation."
- "AI data centers should pay their own way on energy so that households aren’t subsidizing them; and they should generate local jobs and tax revenue." Alongside this, public-private partnerships should accelerate grid expansion, including "a narrow federal authority to accelerate the construction of interregional transmission when it is in the national interest."
- Market failure is the stated justification for intervention: "In normal times, the case for letting markets work on their own is strong", but "industrial policy can play an important role when market forces alone aren’t sufficient", using "research funding, workforce development, market-shaping tools, and targeted regulation."
- Workers should get "a formal way to collaborate with management to make sure AI improves job quality, enhances safety, and respects labor rights", and policy should "set clear limits on harmful uses of AI that could erode job quality by intensifying workloads, narrowing autonomy, or undermining fair scheduling and pay."
- Tax policy should adapt so programmes like Social Security, Medicaid and SNAP remain funded: policymakers "could rebalance the tax base by increasing reliance on capital-based revenues—such as higher taxes on capital gains at the top, corporate income, or targeted measures on sustained AI-driven returns—and by exploring new approaches such as taxes related to automated labor."
- A Public Wealth Fund "that provides every citizen—including those not invested in financial markets—with a stake in AI-driven economic growth", seeded jointly by policymakers and AI companies, with returns that "could be distributed directly to citizens."
- Safety nets should scale automatically: a package of temporary expanded benefits "that activates automatically when these metrics exceed pre-defined thresholds" and phases out as conditions stabilise, avoiding "a permanent expansion of programs."
- Strengthen CAISI "to develop auditing standards for frontier AI risks in coordination with national security agencies" and build "a competitive market of auditors and evaluators" through procurement, advance-purchase commitments, insurance frameworks and standards-setting.
- Stronger controls, "including pre- and post-deployment audits using the standards developed in advance", may eventually be needed for models that could materially advance CBRN or cyber risks — but should apply "only to a small number of companies and the most advanced models, preserving a vibrant ecosystem of less powerful systems and the startups building on them."
- Incident reporting: "Establish a mechanism for companies to share information about incidents, misuse, and near-misses with a designated public authority", including cases where "models exhibited concerning internal reasoning, unexpected capabilities, or other warning signals—even if safeguards ultimately prevented harm."
- Frontier AI companies "should adopt governance structures that embed public-interest accountability into decision-making, such as Public Benefit Corporations with mission-aligned governance", and harden systems "so no individual or internal faction can quietly use AI systems to concentrate power."
- Government use of AI needs "clear rules for how governments can and cannot use AI, with especially high standards for reliability, alignment, and safety", which "should be codified in law"; FOIA and federal-records rules should be modernised so AI-interaction and agentic action logs can count as federal records.
- Internationally, expand CAISI's role and build "a global network of AI Institutes", with policymakers ensuring companies can share safety information "without running afoul of antitrust or competition constraints, using clear safe harbors and narrowly scoped information-sharing rules."
About this document
A 13-page PDF of roughly 5,500 words, produced from Google Docs and issued in OpenAI's corporate name with no individual author or signatory; the landing page lists the author simply as "OpenAI" under Global Affairs. The cover (page 1) carries the title and "April 2026". It opens with a three-page framing essay headed "Let’s Talk" that sets three aims (share prosperity broadly, mitigate risks, democratise access and agency) and argues "The Case for a New Industrial Policy" by analogy to the Progressive Era and the New Deal. The body is two numbered sections. "1. Building an Open Economy" (pages 5-8) sets out twelve bolded proposals: worker perspectives, AI-first entrepreneurs, a Right to AI, modernising the tax base, a Public Wealth Fund, accelerated grid expansion, efficiency dividends, adaptive safety nets, portable benefits, pathways into human-centred work, and accelerated scientific discovery. "2. Building a Resilient Society" (pages 9-12) sets out nine: safety systems for emerging risks, an AI trust stack, auditing regimes, model-containment playbooks, mission-aligned corporate governance, guardrails for government use, mechanisms for public input, incident reporting, and international information-sharing. A one-page close, "Starting the Conversation", announces a feedback inbox, fellowships and research grants of up to $100,000 plus up to $1 million in API credits, and discussions at a new OpenAI Workshop in Washington, DC. There are no footnotes, citations or statutory references; the only laws named are the EU AI Act and unnamed "US state-based regulation", mentioned as where upstream safeguards have already been codified.
How this sits against AI law
Each stance compared with what EU and US instruments actually require. Where no instrument addresses a theme, that gap is shown rather than hidden.
Pre- and post-deployment audits for a narrow set of frontier models
CAISI should develop auditing standards for frontier AI risks and government should build a competitive market of auditors. Models that could materially advance CBRN or cyber risks may come to require pre- and post-deployment audits against those standards, applied only to a small number of companies and the most advanced models.
Article 55(1)(a) obliges providers of general-purpose models with systemic risk to evaluate their models, including adversarial testing, but leaves it to the provider whether that testing is internal or external; independent evaluation arises only through the voluntary GPAI Code of Practice or when the AI Office itself appoints experts under Article 92.
SB 53 requires a large frontier developer's published frontier AI framework to describe its use of third parties to assess catastrophic risk, but it does not make an external audit a condition of deployment, so a mandatory audit regime would go beyond it.
Incident and near-miss reporting to a public authority
Companies should share information about incidents, misuse and near-misses with a designated public authority, under a system that emphasises learning and prevention over punishment. Near-misses should include models showing concerning internal reasoning or unexpected capabilities even where safeguards prevented harm.
Article 55(1)(c) requires systemic-risk providers to track, document and report serious incidents and possible corrective measures to the AI Office without undue delay, but its trigger is an incident that occurred; there is no duty to report near-misses or warning signals where safeguards held.
SB 53 requires frontier developers to report critical safety incidents to California's Office of Emergency Services within 15 days, or within 24 hours where there is an imminent risk of death or serious physical injury; its reportable events are defined incidents tied to catastrophic risk, not a general near-miss channel.
A national evaluation body at the centre of international coordination
CAISI should be expanded into a trusted technical body for evaluating frontier systems and seed a global network of AI institutes that share evaluation results and alignment findings, with antitrust safe harbours so companies can share safety information.
The Act already gives the EU a central public evaluator for frontier models: the AI Office supervises general-purpose models, can request information and conduct evaluations under Articles 91-92, and is advised by a scientific panel of independent experts. It contains no antitrust safe harbour for inter-company safety information-sharing.
The Action Plan gives CAISI the lead on evaluating frontier AI systems for national security risks in partnership with frontier developers. That role rests on voluntary cooperation: no federal instrument obliges a developer to submit a model to CAISI.
Worker voice and limits on harmful workplace AI
Workers should have a formal way to collaborate with management so AI improves job quality and safety and respects labour rights, with clear limits on uses that intensify workloads, narrow autonomy or undermine fair scheduling and pay.
Article 5(1)(f) prohibits emotion-recognition systems in the workplace except for medical or safety reasons, and AI used in recruitment and worker management is classed as high-risk under Annex III. The duty on deployers to inform workers' representatives (Article 26(7)) belongs to the high-risk regime whose application the 2026 amending regulation deferred. Nothing in the Act gives workers a say in deployment decisions.
The Action Plan's worker measures are about skills, retraining and research on AI's labour-market effects. It gives workers no role in deciding how AI is deployed and sets no limit on workplace uses.
Sharing AI gains through the tax base and a Public Wealth Fund
Tax policy should shift toward capital-based revenues and possibly taxes related to automated labour, so that programmes like Social Security and Medicaid stay funded. A Public Wealth Fund seeded jointly by policymakers and AI companies would give every citizen a direct stake in AI-driven growth.
The AI Act regulates products and models, not fiscal policy, and no EU instrument in the corpus taxes AI returns or creates a public stake in AI firms.
No enacted US law does either. Two federal bills would: S. 4825, the American A.I. Sovereign Wealth Fund Act, proposes a fund built from a one-time equity levy on major AI firms, and H.R. 10044, the AI Tax and Work Protection Act, proposes an excise tax on AI developers and deployers. Both are only proposals.
Data-centre energy costs and grid expansion
AI data centres should pay their own way on energy so households are not subsidising them, and should generate local jobs and tax revenue. Public-private partnerships, along with a narrow federal authority for interregional transmission, should speed up grid expansion.
The AI Act touches energy only through documentation and standardisation: providers of general-purpose models record known or estimated energy consumption, and the Commission is to seek standards on energy-efficient AI. Who pays for grid capacity, and how grids expand, is outside its scope.
EO 14318 fast-tracks federal environmental and permitting reviews for data-centre projects above 100 megawatts and streamlines related transmission approvals, matching the acceleration half of OpenAI's ask. It contains no rule that data centres bear their own energy costs.
Statutory guardrails on government use of AI
Policymakers should set clear rules, codified in law, for how governments can and cannot use AI, with especially high standards for reliability, alignment and safety. FOIA and records law should be modernised so AI-interaction and agentic action logs can be retained and reviewed.
The EU has already put rules on public-sector AI into statute. Article 5 prohibits social scoring and restricts real-time remote biometric identification by law enforcement. Public bodies deploying high-risk systems will also have to carry out a fundamental-rights impact assessment under Article 27 once the high-risk regime applies.
M-25-21 sets minimum risk-management practices for high-impact AI used by federal agencies, including pre-deployment testing and impact assessments. It is an executive memorandum rather than statute, and it does not address AI logs as federal records.
Most of this agenda sits outside the terrain AI regulation currently occupies. The EU AI Act is product-safety law: it has something to say about frontier-model evaluation, serious-incident reporting and public-sector use, but nothing about tax, public wealth funds, safety nets or who pays for grid capacity, so on the economic half of the document there is simply no EU or US counterpart to align with, only proposed federal bills. Where the document does meet existing law, it asks for more than the law requires on the frontier-safety side — mandatory pre- and post-deployment audits and near-miss reporting go beyond both the AI Act's self-run evaluation duties and California SB 53's disclosure regime — while insisting that such controls reach only "a small number of companies". In the US the pull runs in two directions: its infrastructure asks match federal executive policy that speeds data-centre permitting, while its call for statutory limits on government AI and a mandatory audit regime goes past a federal posture built on executive memoranda and voluntary engagement.
Source
https://cdn.openai.com/pdf/561e7512-253e-424b-9734-ef4098440601/Industrial%20Policy%20for%20the%20Intelligence%20Age.pdf- Date on the page:
- April 2026
- Source checked:
- opened and confirmed on 2026-09-29