Policy Recommendations

In February 2025, delegations from roughly sixty countries gathered under the glass roof of the Grand Palais in Paris for an AI Action Summit — the third in a series that had begun with genuine optimism at Bletchley Park sixteen months earlier. The summit produced a declaration committing signatories to make artificial intelligence "open, inclusive, transparent, ethical, safe, secure and trustworthy." China signed it. India signed it. Every member state of the European Union signed it. The two governments that host most of the world's frontier AI laboratories — the United States and the United Kingdom — did not. The American vice president used his time at the podium not to discuss safety but to warn Europe against "excessive regulation" that might strangle a world-changing industry in its cradle.

That refusal is the right place to begin a chapter of policy recommendations, because it dramatizes the problem every recommendation here has to survive. There is no shortage of sensible things to propose. Economists, ethicists, technologists, and jurists have produced a large and surprisingly convergent literature on what good AI governance would look like, and much of what follows will feel, to a reader who has been paying attention, almost obvious. The difficulty was never knowing what to do. The difficulty is the distance between the governance we can describe on paper and the governance a fractured, competitive, unevenly capable world can actually deliver. A chapter that lists twenty excellent recommendations and stops there has done the easy ninety percent and skipped the hard ten.

So this chapter tries to do two things at once. It sets out an agenda across five domains — economic policy, rights, democratic governance, safety, and international coordination — grounded in the analysis of the preceding chapters. And then it turns that agenda on itself, asking which parts are likely to happen, which are wishful, and whether the fine principles in its own preamble actually survive contact with its own proposals. The second task matters as much as the first. An honest reader deserves to know not just what would help, but which of the things that would help we are actually capable of doing.

Eight Principles, Offered as a Test

Any coherent governance agenda needs a value framework — both to guide individual choices and to arbitrate when good things conflict. Eight principles underwrite the recommendations that follow, and it is worth stating them plainly before examining whether the recommendations honor them.

The first is human agency preservation: AI should extend human capability, not quietly displace human judgment in decisions that matter. The second is broad benefit distribution: the productivity gains AI generates should reach the population widely rather than pooling among those who own the capital and the models. The third and fourth address uncertainty. Reversibility holds that, because we understand AI's trajectory so poorly, policy should preserve the ability to change course and avoid lock-in. Proportionate precaution holds that where a harm could be catastrophic, the burden of proof shifts: a system with the potential for severe damage should demonstrate its safety before deployment, not after. The last four concern structure. Democratic governance insists that consequential choices be made through legitimate public processes rather than by technical or commercial elites alone. Global cooperation recognizes that AI's challenges cross borders and that race-to-the-bottom dynamics can only be countered together. Transparency and accountability demands that consequential systems be legible enough to oversee, with someone identifiable to hold responsible when they fail. And equity requires that policy narrow existing inequalities rather than harden them.

These are not controversial. They cross political and cultural lines, and a reader in Nairobi, Berlin, or São Paulo could endorse all eight without straining. That very breadth is the warning sign. Principles this agreeable are easy to invoke in a preamble and easy to lose in the details. So treat this list not as decoration but as a test — a rubric this chapter will apply to its own recommendations before it ends, to see which principles do real work and which merely make an appearance.

Economic Policy: Letting the Obstacles Do Their Job

The economic recommendations follow directly from the book's central claim, argued at length in earlier chapters: that AI's distributional consequences depend far more on policy than on the technology itself. Four interventions matter most, and each is more interesting once its obstacles are allowed to reshape it rather than being waved past.

The flagship is a universal AI dividend — a basic income, or something functionally close, funded by taxing the productivity that automation captures. The evidence base here is genuinely encouraging, and it is worth being precise about what we actually know. Alaska has paid every resident an annual dividend from its oil fund since 1982, and a careful study of that program found no significant reduction in overall employment (Jones & Marinescu, 2018). Finland's two-year experiment paying €560 a month to 2,000 unemployed people produced small positive employment effects and large improvements in wellbeing, trust, and mental health (Kela, 2020). Stockton, California's SEED pilot found that full-time employment among recipients of $500 a month rose faster than in a control group (West & Castro, 2021). The most rigorous test to date, the OpenResearch study funded partly by Sam Altman, gave 1,000 low-income Americans $1,000 a month for three years: recipients worked modestly less — on the order of a little over an hour a week — spent more on food, rent, and helping others, and showed greater interest in education and starting businesses (OpenResearch, 2024). This is what we know. What we do not know is whether any of it scales. A pilot funded from outside does not have to solve the two problems that a national program must: how to raise the money without the tax base fleeing, and how to sustain the political coalition across decades.

Those two obstacles should change the shape of the recommendation, not merely accompany it. A dividend financed by taxing corporate profits invites exactly the offshoring and profit-shifting that multinationals have perfected; a dividend financed by taxing things that cannot easily leave — domestic compute consumption, data used to serve domestic customers, consumption itself — is harder to evade. The Alaska model suggests a second design lesson: a modest, universal, visibly-shared payment indexed to a common resource has proven far more politically durable than a means-tested transfer, precisely because everyone receives it and everyone therefore defends it. The honest version of this recommendation is not "implement UBI" but "build a modest, hard-to-evade, universal dividend now, while the sums are small, so that the machinery and the political constituency exist before the automation wave that would require them."

The same discipline applies to retraining and transition support, where honesty is uncomfortable. The historical record of retraining programs is sobering: evaluations of the United States' Trade Adjustment Assistance and comparable schemes have repeatedly found that displaced workers rarely recover their former earnings, and that reskilling alone is a weak instrument against structural displacement. Allowing that obstacle to reshape the recommendation pushes it away from faith in reskilling and toward income security during transition — wage insurance that tops up the pay of workers who move to lower-paying roles, and the comprehensive, well-funded active labor market policies that the Nordic countries, alone among rich economies, have shown can actually work. The lesson is not that transition support is futile but that thin, underfunded versions are, and pretending otherwise sets workers up to be blamed for the failure of programs that were never resourced to succeed.

The remaining two economic recommendations rest on firmer ground. Pro-competition policy must confront the reality that AI markets tend toward concentration through data network effects and the staggering capital cost of frontier training — which argues for interoperability and data-portability mandates, public compute infrastructure available to smaller entrants, and antitrust attention to algorithmic price coordination, which can reproduce a cartel's effects with no cartel ever meeting. And stakeholder governance — putting workers and public-interest representatives on the boards of the largest AI firms — has a real precedent in Germany's codetermination laws, which economic research has found neither wrecked competitiveness nor much raised wages, but did modestly shift firms toward longer-horizon investment (Jäger, Schoefer & Heining, 2021). That is a realistic claim about a realistic tool: not a transformation, but a moderation of the most extractive impulses of purely shareholder-driven management.

Rights: The Problem With "A Human Reviews It"

The rights recommendations translate the book's chapters on surveillance, discrimination, and autonomy into enforceable protections: a right to human decision-making in consequential domains, comprehensive data rights, algorithmic transparency and auditing, and anti-discrimination rules that reach disparate impact, not just intent. Three of these are relatively tractable and already partly underway — the EU's GDPR established a data-rights template that California's CCPA showed a large jurisdiction can adopt independently, and the EU AI Act now requires disclosure, documentation, and auditing for high-risk systems (EU AI Act, 2024). One of them has a hole in the middle, and the hole deserves the chapter's full attention.

The right to human decision-making promises that in healthcare, credit, employment, criminal justice, and education, a person can insist that a human being make — or at least meaningfully review — a decision that affects them. The legal text of the recommendation is careful to say the review must be "substantive rather than nominal." But the book's own analysis of cross-cutting themes explained, in detail, exactly why that promise tends to hollow out. As an AI system becomes more accurate than the humans nominally supervising it, deferring to it stops being laziness and starts being rational. The reviewer who overrides a system that is right ninety-eight percent of the time will, on average, make things worse and will be blamed when the override goes wrong. Automation bias is one of the most robust findings in the human-factors literature, documented for decades in aviation and medicine: people under time pressure, facing a confident machine recommendation, approve it. A "human in the loop" who rubber-stamps is not oversight. It is a liability shield with a pulse.

A recommendation that ignores this mechanism is writing a right it knows will be defeated in practice. The more defensible version accepts the diagnosis and redesigns around it. If genuine human deliberation cannot be conjured by statute, then the enforceable thing is not the deliberation but its preconditions and its traces. That means giving the reviewer real authority, real time, and — crucially — protection from liability for justified overrides, so that disagreeing with the machine is not career-ending. It means auditing override rates as a diagnostic: a reviewer who never once departs from the algorithm is not evidence of a good algorithm but of a dead review. And it means shifting the legal center of gravity from "a human decided" toward two things that can actually be checked — a right to a genuine explanation, and a right to contest the decision before a human with the power to reverse it. In some domains the honest conclusion is starker still: where meaningful review is genuinely unattainable, the right should be a prohibition on fully automated decision-making, not a requirement for a review everyone knows is theater. The European Court of Justice moved in roughly this direction in 2023 when it ruled that automated credit scoring falls under the GDPR's constraints on solely-automated decisions (CJEU, SCHUFA, 2023) — treating the automated decision itself, not the fiction of human sign-off, as the thing to regulate.

Governance and Safety: The Institutional Scaffolding

Two domains supply the machinery that makes everything else enforceable. On the governance side, the recommendations call for dedicated national AI agencies with technical staff and real authority, public deliberation through citizens' assemblies and consequential consultation, algorithmic impact assessments modeled on environmental review, and sunset clauses that force regulations to be renewed on evidence rather than persist by inertia. The first of these is no longer hypothetical: the EU has stood up an AI Office, and a string of countries created AI safety institutes in the wake of the Bletchley summit. The FDA, FAA, and FCC are proof that specialized technical regulators can hold their own against powerful industries over the long run — imperfectly, and always fighting the gravitational pull of regulatory capture, but really. Sunset clauses deserve a specific note, because they are the one recommendation that directly operationalizes the reversibility principle: a five-to-seven-year renewal cycle is frequent enough to catch a rule that technological change has made obsolete, and infrequent enough to give the regulated some certainty to plan against.

On the safety side, the recommendations borrow, deliberately, from industries that learned to manage catastrophic risk: safety standards and certification calibrated to a system's risk level, mandatory incident reporting modeled on aviation's remarkably successful practice of protecting good-faith reporters so that the whole industry learns from each near-miss, capability limitations that flatly prohibit the highest-risk applications — autonomous lethal weapons, AI assistance in designing bioweapons, uncontrolled self-replicating systems — and resilience requirements ensuring that when an AI-dependent power grid or hospital network fails, human-operated backups still exist and the humans still remember how to run them. This last point is easy to underrate. A grid made more efficient by AI optimization but incapable of functioning when the AI fails has not become more resilient; it has traded a familiar fragility for an unfamiliar one, and unfamiliar fragilities are the ones that produce disasters.

Incident reporting and safety certification are, of everything in this chapter, among the most likely to arrive, precisely because they ask industry for something it can live with and even benefit from. Capability limitation is a different animal, because it depends on the one thing this book has consistently described as scarce: international coordination.

International Coordination: Where the Chapter Must Not Lie to Itself

The four international recommendations — a global AI governance framework on the model of the IAEA or WHO, an international safety-research consortium, capability-control agreements, and development support and technology transfer to the global South — are simultaneously the most important recommendations in the chapter and the least likely to be realized at anything near the needed scale. The book's earlier sections on geopolitics made the case at length that meaningful coordination between the United States and China on a technology both regard as decisive is extraordinarily hard. A recommendations chapter that then breezily asserts such coordination is "both necessary and achievable" is not being brave; it is being evasive.

The honest move is to distinguish the coordination that history says is achievable from the coordination that history says is not. Look at what worked. The Montreal Protocol phased out ozone-depleting chemicals and is, by every scientific assessment, healing the ozone layer — but it worked because the harm was clear, the substances were made by a handful of firms, substitutes existed, the economic sacrifice was modest, and compliance was verifiable. Nuclear test-ban and non-proliferation regimes have held, imperfectly, because the relevant materials and detonations are physically detectable. Now look at what failed, or barely moved. The Kyoto and Paris climate agreements asked for broad, economically painful, hard-to-verify sacrifice against a diffuse harm, and global emissions kept rising. AI capability sits closer to the climate case than the ozone case: it is dual-use, economically central, and its most dangerous capabilities are lines of weights on a server, not plumes a satellite can photograph.

That comparison should reshape the international agenda rather than merely qualifying it. The coordination worth pursuing first is the coordination that shares the features of the successes — narrow, verifiable, and grounded in genuine mutual interest. A ban on fully autonomous lethal weapons, a shared incident-reporting database, and above all a jointly funded safety-research consortium fit that description, because a breakthrough in alignment or interpretability makes every nation safer at no one's competitive expense. CERN, the Human Genome Project, and decades of international climate science show that pre-competitive scientific collaboration across deep geopolitical divides is achievable when the scope is clearly shared and the governance is seen as fair. The comprehensive "IAEA for AI" with binding enforcement over frontier development is a worthy aspiration and should be named as one, but a chapter that presents it as a near-term deliverable is selling the reader a coordination it has no reason to expect. And development support — the recommendation most directly tied to the equity principle — is worth pursuing on its merits, but it is precisely the kind of costly, altruistic, low-reciprocity commitment that the historical record suggests wealthy nations sustain least reliably. Naming that is not defeatism. It is the difference between a plan and a wish.

Do the Principles Survive?

Now the promised test. The chapter opened with eight principles. Applying them to the twenty recommendations honestly, they do not come out even.

Three principles — distribution, transparency, precaution — are genuinely load-bearing; the recommendations that serve them are concrete, and several are already being enacted somewhere. Human agency is nominally served by the right to human decision-making, but that is exactly the recommendation the book's own analysis predicts will hollow into rubber-stamping unless it is redesigned around contestability. The bottom four are where the framing outruns the substance. Democratic governance is invoked as foundational, yet the recommendations meant to deliver it — deliberation, consultation — are the softest and least enforceable in the entire chapter. Reversibility appears prominently in the preamble and is operationalized by exactly one mechanism, sunset clauses, which sits in open tension with the chapter's own flagship: a universal basic income, once established, is among the most politically irreversible programs a state can create. And equity across nations, though it animates the international section's rhetoric, is carried by the recommendations least likely to be funded and sustained.

This is the honest answer to the question the chapter is built around. The eight principles are not equally operationalized. Some do real work; others appear in the framing and thin out in the substance — and one, reversibility, is contradicted by the chapter's own centerpiece. Naming that mismatch is not an admission of failure. It is what distinguishes a set of recommendations that has examined itself from a wish list that has not.

The Sequencing Problem

The previous version of this agenda proposed a three-phase rollout, and the structure is sound enough to keep — provided its internal dependencies are examined rather than assumed away.

Two dependencies inside this table create trouble. The first is the UBI evidence crunch. Phase 3 promises to "scale UBI based on pilot evidence," but the pilots do not begin until Phase 2 — meaning years two through five. A three-year pilot launched even midway through Phase 2 does not report until year six or seven, and evaluating results, then building the political coalition to act on them, easily consumes the rest of the decade. The evidence that is supposed to justify Phase 3's central economic commitment barely exists when Phase 3 begins. The realistic repair is the one the economic section already argued for: start a modest universal dividend in Phase 1, small enough to be affordable and reversible, so that a decade of real data and a real constituency exist by the time scaling is on the table. The second dependency is coordination time. Phase 3's international institution rests on negotiations opened in Phase 1 and a framework concluded in Phase 2 — but the WTO, the International Criminal Court, and the climate regime each took well over a decade to move from first talks to functioning body. Compressing that into five years, for a technology two great powers regard as strategically decisive, is the sequencing equivalent of the coordination optimism the chapter just cautioned against. The phasing is useful as a statement of dependencies and priorities. It is misleading as a calendar.

What "Partial Implementation" Can and Cannot Do

The agenda's animating consolation is the claim that "partial implementation shifts trajectories; comprehensive implementation transforms them." It is a comforting line, and it is worth asking whether the evidence supports it or whether it is doing emotional work — softening the chapter's own admission that much of this will not be implemented at the scale required.

The answer depends entirely on the shape of the harm. For harms that are cumulative and divisible, the claim holds up well. Tobacco control's Framework Convention was never universally or fully enforced, yet every increment — each tax rise, each advertising ban, each smoke-free workplace — measurably reduced smoking and saved lives. A basic income that reaches some workers, an anti-discrimination rule enforced in some sectors, a transparency mandate that covers some systems: each of these genuinely bends the curve for the people it touches, because the harm it addresses is the sum of many separable instances. Here, partial really does shift the trajectory.

For harms that are threshold and indivisible, the claim collapses. If the danger is that a single actor develops an AI system capable of designing a pandemic pathogen, then a capability-control agreement signed by every nation except one has not reduced the risk by the fraction that signed — it has failed, because the harm requires only one defector. Montreal worked because near-universal compliance was achievable and the last few holdouts could not undo the whole; a bioweapon regime does not have that property. So the slogan is half true, and the half it omits is the dangerous one. Partial implementation is a real gain against the divisible harms that dominate the economic and rights domains, and close to worthless against the indivisible catastrophic harms that dominate the safety and international domains — which are, not coincidentally, exactly the domains where full coordination is least achievable. The reader is owed both halves.

If You Could Only Do a Few

Comprehensive implementation is not politically available, and a chapter that implicitly assumes the full agenda is equally achievable has abandoned the reader at the moment they most need guidance. So which recommendations should go first — and which should we expect to wait?

The pattern that emerges is not random, and it is the most useful thing this chapter can leave the reader with. The recommendations most likely to be adopted share three features: they ask industry for something it can absorb, they build on templates that already exist, and they address divisible harms where partial coverage still helps. Transparency, incident reporting, and technical regulators fit all three — and, tellingly, are already appearing in the EU AI Act and the network of national safety institutes. The recommendations facing the highest barriers share the opposite features: they demand large fiscal or sovereign sacrifice, they lack precedent at the needed scale, and several address indivisible harms where partial adoption fails. A universal dividend at national scale and enforceable capability-control agreements sit at that hard end — the two most consequential proposals in the chapter and the two least likely to arrive on time.

If forced to prioritize, then, the strategic logic is to bank the achievable institutional and transparency gains first — not because they are the most important, but because they build the state capacity, the data, and the public legitimacy that the harder redistributive and international measures will require later. Establish the agencies that will one day have to enforce a capability agreement. Start the modest dividend that will one day have to scale. Build the auditor corps before the audits are mandatory everywhere. The sequencing that makes sense is not the one that does the most valuable things first, but the one that makes the most valuable things possible.

Key Takeaways

AI governance is not one problem with one solution but a cluster of overlapping challenges across economic distribution, individual rights, democratic accountability, safety, and international coordination. This chapter set out twenty recommendations across those five domains and then held them to the eight principles it opened with. Three conclusions matter most.

First, the principles are not equally honored. Broad benefit distribution, transparency, and proportionate precaution are richly operationalized by concrete, partly-enacted recommendations. Democratic deliberation, reversibility, global cooperation, and cross-national equity appear prominently in the framing but thin out in the enforceable substance — and reversibility is actively contradicted by the chapter's own flagship, a basic income that is, by design, extremely hard to reverse.

Second, feasibility is not background noise; it should reshape the recommendations. Letting obstacles do their work turns "implement UBI" into "start a modest, hard-to-evade dividend now and scale it on evidence," turns "reskill displaced workers" into "insure their wages while they transition," turns "a human reviews consequential decisions" into "the decision is explained, contestable, and reversible by an accountable person," and turns "build an IAEA for AI" into "start with the narrow, verifiable, mutual-interest agreements that history says actually hold."

Third, the comforting claim that partial implementation shifts trajectories is true for the divisible harms of the economic and rights domains and close to false for the indivisible catastrophic harms of the safety and international domains — which are precisely where full coordination is hardest to get. The prudent path is to secure the achievable institutional and transparency gains first, not because they matter most, but because they build the capacity, evidence, and legitimacy that the harder measures depend on. The central choice is not whether AI will be governed but whether that governance will be proactive and legitimate, or reactive and written, by default, in the interests of those with the most to gain from no rules at all.

Sources

• International AI Safety Report 2026
• EU Artificial Intelligence Act — official text and implementation timeline
• OpenResearch — Unconditional Cash Study
• Alaska Permanent Fund Dividend Division
• Kela — Results of Finland's basic income experiment
• How 2026 Could Decide the Future of AI | Council on Foreign Relations
• How the world can build global AI governance | World Economic Forum
• Eight ways AI will shape geopolitics in 2026 | Atlantic Council

Last updated: 2026-09-23