Start with the variables, not the movie plot
There is no observation of a post-AGI society. “After AGI” is therefore not a factual period we can describe in the way we describe an election, recession, or product release. It is a family of scenarios built from assumptions about capability, cost, access, ownership, safety, and political response. A useful scenario makes those assumptions visible. A bad one hides them behind a vivid story.
Even the technical starting point is uncertain. The International AI Safety Report 2026 describes rapid improvement alongside jagged capabilities, an evaluation gap between tests and real-world performance, and several plausible development paths through 2030. It does not endorse one arrival date or one inevitable trajectory. That uncertainty should carry through to social analysis.
Four variables do most of the work:
- Speed: does broadly transformative capability diffuse over years, or arrive faster than employers, regulators, schools, and households can adjust?
- Access: is capability available to millions of organizations and individuals, or controlled through a few providers and governments?
- Ownership: who owns the models, chips, energy infrastructure, data centers, complementary businesses, and resulting income?
- Control: do institutions retain the ability to inspect, restrict, replace, and appeal AI-mediated decisions?
The scenarios below are not predictions with assigned probabilities. They are stress tests. Reality could combine features from all four, differ by country, or move from one to another.
Scenario one: managed diffusion
In managed diffusion, capability improves quickly but not instantly. Systems automate larger bundles of work, yet deployment remains limited by integration costs, reliability, law, physical infrastructure, and organizational inertia. Governments expand transition support, firms redesign jobs, and safety requirements mature before any one system becomes indispensable.
This is the scenario most continuous with present evidence. The International Labour Organization’s 2025 exposure index finds that one in four workers is in an occupation with some generative-AI exposure, but argues that job transformation is more likely than wholesale replacement because occupations contain tasks that still require human input. Exposure is not the same as adoption, and adoption is not the same as eliminating a job.
Managed diffusion does not guarantee fairness. Entry-level pathways can shrink before whole occupations disappear. Productivity gains can flow to owners while wages stagnate. Regions with weak digital infrastructure can fall behind. The difference is that institutions have time to observe these effects and respond through training, bargaining, taxation, competition policy, public investment, and changes to working time.
The household strategy in this world is adaptation rather than retreat: build skills that complement new tools, reduce dependence on one employer, preserve human networks, and participate in the institutions deciding how systems are deployed.
Scenario two: concentrated acceleration
In concentrated acceleration, a small number of firms achieve a large capability lead and keep access controlled. Their systems become essential inputs to research, software, finance, logistics, media, and public administration. Economic output rises, but bargaining power shifts toward the owners of models and compute.
Concentration is not hypothetical even if AGI is. A US Federal Trade Commission staff study documented equity and revenue-sharing rights, consultation or control rights, exclusivity, cloud-spending commitments, and information sharing in three partnerships linking Alphabet, Amazon, and Microsoft with Anthropic or OpenAI. The sample was limited to those partnerships and reflected information available to staff through September 2024 plus public material through January 2025; it should not be generalized to every cloud–developer relationship. The report identified potential switching costs and dependencies rather than adjudicating an antitrust violation.
In this scenario, material abundance can coexist with political fragility. Governments may rely on providers they cannot readily replace. Workers may receive cheaper services while losing income and negotiating power. Citizens may formally retain rights but struggle to contest decisions made through proprietary systems. A company need not intend to govern society for dependency to give it governance-like power.
The decisive policies are not emergency stockpiles. They are ownership, interoperability, taxation, antitrust, procurement, audit access, public options, and limits on delegating sovereign functions. The central question is whether the public can change provider or policy without disabling systems on which ordinary life depends.
Scenario three: open proliferation
In open proliferation, high capability becomes widely accessible through open weights, cheap inference, copied systems, or intense competition. Innovation spreads quickly. Small firms, researchers, communities, and lower-income countries gain tools that would otherwise remain behind expensive APIs. No single provider becomes an unavoidable gatekeeper.
The benefit is distributed agency. The risk is distributed dangerous capability. Once weights are widely mirrored, the original publisher cannot reliably recall every copy, and safeguards can be modified. The US National Telecommunications and Information Administration’s July 2024 open-model report concluded that the evidence available then did not justify blanket restrictions, while recommending monitoring and capacity to respond if future models changed the balance. That dated, conditional position is more useful than treating “open” as automatically safe or automatically reckless.
An open-proliferation society would need defenses outside the model: secure biological supply chains, hardened critical infrastructure, fraud-resistant identity systems, rapid incident sharing, and law-enforcement capacity that does not depend on suppressing general knowledge. It would also need governance that distinguishes a small research model from a system that materially enables cyber operations, weapons development, or autonomous replication.
This scenario gives more people access to productive capacity, but access alone does not settle distribution. People still need compute, energy, expertise, customers, and legal permission to turn a model into durable income.
Scenario four: capability shock and institutional lag
In a capability shock, reliable systems begin completing economically important work faster than organizations can redesign roles or governments can replace lost income and tax revenue. The defining feature is not intelligence in the abstract. It is the gap between technical change and institutional response.
That gap can create contradictory conditions: high productive capacity alongside unemployment, valuable automated services alongside collapsing business models, and governments asked to provide more support while conventional payroll-tax bases weaken. It can also increase pressure to delegate decisions because human review looks slow precisely when oversight is most important.
The worst outcomes are not automatic. Existing systems such as unemployment insurance, deposit insurance, automatic fiscal stabilizers, collective bargaining, emergency appropriations, and public procurement provide starting infrastructure. But programs designed for temporary recessions may not fit a persistent change in the value of labor. Decisions about income, ownership, and public revenue would need to move from academic debate to operational policy quickly.
Households cannot privately insure against a society-wide capability shock. Personal savings buy time, but the durable response is collective: institutions that can distribute purchasing power, keep essential services operating, prevent predatory concentration, and revise policy as evidence arrives.
Signals that distinguish the scenarios
Watch measurable conditions rather than AGI branding:
- the reliability and cost of completing whole workflows, not isolated benchmark questions;
- whether adoption produces reduced hiring, lower hours, lower wages, or actual layoffs;
- market concentration in cloud, chips, models, energy, and distribution;
- the share of leading capability available as open weights, controlled APIs, or internal systems;
- whether public agencies retain tested non-AI operating modes and provider portability;
- how productivity gains are divided among prices, wages, profits, taxes, and shorter working time;
- the time between a demonstrated dangerous capability and an enforceable mitigation.
These signals can move in different directions. Cheap open models could reduce corporate concentration while increasing misuse risk. Strong licensing could improve oversight while entrenching incumbents. Rapid productivity growth could fund generous transition policy or increase inequality, depending on ownership and political choices.
Normative goal: robustness rather than certainty
A robust policy performs tolerably across several scenarios. Phishing-resistant identity helps whether systems are centralized or open. Portable benefits help whether jobs disappear or simply churn faster. Interoperability reduces dependency without requiring a forecast about which provider wins. Incident reporting improves learning under both voluntary frameworks and binding regulation. Public investment in education, energy, health, and administrative capacity creates value even if capability growth slows.
The point of scenario analysis is not to choose the most dramatic future. It is to identify decisions that become expensive to reverse. Ownership structures, critical-infrastructure dependencies, surveillance systems, and the loss of human expertise can lock in before anyone agrees that AGI has arrived. Those are the decisions worth making deliberately now.