Artificial general intelligence
Systems able to match or exceed skilled human performance across most economically relevant cognitive tasks at competitive cost. This working definition does not imply that every job is automated.
KNOWLEDGE & FORESIGHT / WORKING FRAMEWORK
Prepare for possibilities.
Protect human agency.
There is no single future of society with advanced AI. Outcomes depend on how systems develop, who controls them and the institutions people build. Our framework connects uncertainty to decisions that emerging economies can act on.
01 / Definitions
Start with clear terms.
AGI and ASI have competing definitions. We use the following working meanings for planning, rather than claiming a universal benchmark.
Artificial superintelligence
Hypothetical systems that substantially surpass the best human performance across nearly all cognitive domains. Capability alone does not establish safe behaviour or consciousness.
We track capability, reliability, autonomy and real-world adoption separately. We do not assume that current methods will necessarily lead to AGI, or that progress will be continuous. Levels of AGI ↗
Five variables shape the transition.
A useful plan should change when the evidence changes. These are the questions we propose to review with partners.
- 01
Arrival & adoption
When would broad capability become technically feasible—and when would institutions actually use it?
- 02
Speed of change
Would capability grow gradually, stall, or accelerate faster than institutions can respond?
- 03
Alignment & control
Can systems act reliably within intended goals, accept corrections and be stopped? Whose goals count, and who can challenge them?
- 04
Ownership & distribution
Who controls models and infrastructure, sets the rules and receives the gains?
- 05
Physical & institutional limits
How do energy, chips, robotics, experiments, finance and regulation constrain deployment?
Six pressures. Practical responses.
The mechanisms below are conditional planning hypotheses. The responses set out proposed FOHI work, not results already achieved. More capable systems could create major benefits as well as risks.
01
Work & livelihoods
If reliable AI substitutes for tasks at lower total cost, some cognitive work could become much cheaper. Employment and wages also depend on demand, new tasks, adoption costs and whether AI complements people. Full occupational replacement does not follow automatically.
Gen-AI: Artificial Intelligence and the Future of Work ↗Proposed response
Map task exposure with workers and employers; test augmentation, accessible training and transition support in partner institutions.
What to monitor
Task costs, job redesign, wage changes and who can use the tools.
02
Ownership & inequality
If owners capture a growing share of value, returns to models, compute and other scarce assets could increase inequality. This is a risk to evaluate, not a fixed outcome: bargaining power, competition and public choices matter.
Gen-AI: Artificial Intelligence and the Future of Work ↗Proposed response
Compare options for benefit sharing and social protection; assess access for smaller firms and institutions alongside distributional effects.
What to monitor
Access costs, ownership concentration and the distribution of gains.
03
The physical economy
Robotics, construction and care involve equipment, safety and physical coordination. Energy, chips, materials and experiments can constrain scale. Intelligence does not remove physical limits; no fixed years-long lag should be assumed.
Energy and AI: energy demand from AI ↗Proposed response
Map critical compute and energy dependencies; assess local infrastructure needs and test continuity plans for partner services.
What to monitor
Power availability, compute access, robotics reliability and supply dependencies.
04
Power & accountability
Control over advanced systems could concentrate economic, informational or strategic influence. Harmful use by people and failures of system control require different safeguards; either can undermine human agency.
International AI Safety Report 2026 ↗Proposed response
Develop vendor accountability, audit rights, human escalation and exit options; examine competition and public-interest oversight.
What to monitor
Vendor dependence, decision authority, audit access and control failures.
05
Information & trust
Cheap synthetic content may make impersonation and manipulation easier. Provenance can help establish origin and edit history, but does not prove truth. Missing credentials do not by themselves make content false.
C2PA and Content Credentials Explainer ↗Proposed response
Combine provenance checks with independent verification, correction procedures, media literacy and privacy-preserving identity practices.
What to monitor
Impersonation incidents, correction speed and evidence quality.
06
Human agency & meaning
Economic productivity is only one part of wellbeing. Human care, art, sport and community may retain value even when machines perform tasks better. We cannot confidently predict meaning, status or social roles at population scale.
Proposed response
Include affected communities in design; preserve meaningful human choice, routes to appeal and opportunities for participation.
What to monitor
People’s ability to choose, contest decisions and participate in institutions.
Rehearse more than one future.
These are illustrative exercises, not a ranking or probability forecast. They can overlap. Each translates a possible development into a review trigger and a proposed institutional response.
01Gradual, broad adoption
Capabilities improve while institutions have time to adjust.
Review trigger
Reliable systems become affordable across more locally important tasks.
Preparation now
Build evaluation capacity, monitor workforce outcomes and improve access for smaller institutions.
02Rapid capability acceleration
Capabilities advance faster than normal review and procurement cycles.
Review trigger
A system gains consequential autonomy or reaches a pre-agreed risk threshold.
Preparation now
Set escalation responsibilities, deployment gates, incident exercises and service fallbacks before expansion.
03Concentrated or fragmented access
A few providers control access, or regions face incompatible rules and unequal infrastructure.
Review trigger
Dependence, exclusion, access restrictions or interoperability problems grow.
Preparation now
Test exit options, portability, accountable procurement and regional cooperation.
04Slower progress or binding constraints
Technical progress, economics or physical bottlenecks delay the expected transition.
Review trigger
Pilots do not show reliable benefit, or infrastructure costs exceed feasible resources.
Preparation now
Keep commitments staged, evaluate demonstrated needs and retain measures useful for today’s AI.
Be explicit about what we know.
Our proposed evidence standard separates observations, conditional mechanisms and speculative scenarios. Every assessment should state its assumptions, limits and what would change the conclusion.
Observed constraints
Computing and physical deployment require infrastructure and energy. Local impacts still need measurement.
Conditional mechanisms
Substitution, ownership and institutional choices can reshape outcomes. Their scale and direction vary by setting.
Deep uncertainty
AGI or ASI feasibility, timing, takeoff speed and long-run control remain uncertain. We assign no numerical scenario probabilities here.
Questions that remain open
Can much more capable systems remain controllable?
Alignment, robust oversight and recovery from loss of control remain central research and governance questions. Limited evaluations cannot establish general safety.International AI Safety Report 2026 ↗
Can institutions adapt to a discontinuous jump?
Plans should include rapid change rather than assuming that years of adjustment will always be available.
How will gains, purpose and status be shared?
Distribution and social meaning depend on political and cultural choices as well as technology. Communities must help define success.
Could AI systems have morally relevant experiences?
Consciousness and moral status remain open questions. We support careful interdisciplinary review without presuming that present systems are conscious.Consciousness in Artificial Intelligence ↗
Turn the framework into accountable work.
For each proposed pilot, record the affected groups, key assumptions, warning signs, responsible human decision-maker and a review date. Publish methods and limitations where appropriate. Revise plans as evidence changes.
See the proposed programmes↗Sources & scope
Primary research and institutional analysis inform this framework. Our scenarios and proposed responses are FOHI’s planning synthesis. These references do not imply endorsement or partnership.
- Levels of AGI ↗Morris et al. · ICML 2024; revised 2025
Definitions and ways to distinguish performance, generality and autonomy.
- Gen-AI: Artificial Intelligence and the Future of Work ↗IMF · 2024
Evidence and modelling on work, complementarity and distribution; not a forecast of universal unemployment.
- Energy and AI: energy demand from AI ↗IEA · 2025
Infrastructure and electricity constraints; projections remain conditional.
- International AI Safety Report 2026 ↗International AI Safety Report · 2026
Evidence on capabilities, misuse, systemic risk and control challenges.
- C2PA and Content Credentials Explainer ↗Coalition for Content Provenance and Authenticity · v2.2
How provenance works and what it can—and cannot—establish.
- 2023 Expert Survey on Progress in AI ↗AI Impacts · 2023 survey
Historical expectations, uncertainty and sensitivity to question framing.
- Consciousness in Artificial Intelligence ↗Butlin et al. · 2023
Scientific approaches to a question that remains unresolved.