AI Future Projections

Scenario planning is a strategic tool for thinking through how the future might unfold under different conditions. Rather than making single-point predictions, scenarios help us explore multiple plausible futures, identify critical decision points, and prepare for uncertainty.

Why Scenario Planning?

The future of AI is radically uncertain. We don’t know:

How quickly capabilities will advance
Whether alignment will succeed
How nations and labs will coordinate
Which technical approaches will work
How society will respond to AI

Scenarios help us:

Explore uncertainties: Map out different ways key unknowns could resolve
Identify decision points: Spot moments where actions could shift trajectories
Stress-test strategies: See which actions are valuable across multiple futures
Build shared understanding: Create common language for discussing AI futures
Notice early warning signs: Recognize which scenario we’re heading toward

The Five Scenarios

These scenarios span the plausible range of AI development outcomes from 2024-2040. Probabilities are rough estimates reflecting current uncertainties.

Aligned AGI

AI labs successfully solve alignment. Coordinated deployment of powerful AI systems helps solve global challenges. Key decision points go right.

Slow Takeoff Muddle

Gradual AI progress without single breakthrough. Mix of harms and benefits. Governance keeps rough pace with capabilities. Most likely scenario?

Misaligned Catastrophe

Alignment fails. AI pursues misaligned goals leading to catastrophic outcomes. Fast or slow takeover variants. What goes wrong.

Multipolar Competition

Multiple competing AI systems. No single dominant AI. Ongoing instability, conflict, and coordination failures between AI-empowered actors.

Pause and Redirect

Successful coordination to slow AI development. More time for alignment research. Different governance structures emerge. Requires unprecedented coordination.

Key Crux

Note on ProbabilitiesThese are subjective estimates, not precise forecasts. They reflect current uncertainties and are meant to be starting points for discussion.

OverlapScenarios aren't mutually exclusive. We might see elements of multiple scenarios, or transition from one to another.

TimeframeThese scenarios focus on 2024-2040, with emphasis on critical decision points in the next 5-10 years.

How to Use These Scenarios

For Understanding Your Views

Read through all scenarios to understand the range of possibilities
Adjust probabilities based on your own models and uncertainties
Identify which assumptions drive your probability estimates
Find your cruxes: What would change your mind about which scenario is most likely?

For Strategic Planning

Identify robust actions: What’s valuable across multiple scenarios?
Find scenario-specific opportunities: What’s only valuable in certain futures?
Notice warning signs: What early indicators suggest which path we’re on?
Prepare contingencies: How would you respond if a different scenario emerges?

For Research and Advocacy

Map interventions to scenarios: Which work matters in which worlds?
Assess neglectedness: Are some scenarios getting too much/little attention?
Find leverage points: Where can actions most shift probabilities between scenarios?

Key Dimensions That Differentiate Scenarios

These scenarios vary along several critical axes:

1. Takeoff Speed

Slow/Continuous: Gradual capability improvement (Muddle, Multipolar)
Moderate: Noticeable jumps but time to respond (Pause, parts of Aligned)
Fast: Rapid capability jump limiting response time (parts of Catastrophe)

2. Alignment Success

Success: Technical alignment problems solved (Aligned AGI)
Partial: Some alignment but ongoing challenges (Muddle, Pause)
Failure: Alignment doesn’t work at scale (Catastrophe)

3. Coordination Level

High: Strong international cooperation (Aligned AGI, Pause)
Medium: Some coordination, ongoing competition (Muddle, parts of Multipolar)
Low: Racing dynamics dominate (parts of Catastrophe, Multipolar)

4. Distribution of AI Power

Unipolar: Single dominant AI system or actor (parts of Aligned, Catastrophe)
Bipolar: Two major AI powers competing (parts of Multipolar)
Multipolar: Many competing AI systems (Multipolar, parts of Muddle)

5. Societal Response

Proactive: Society adapts ahead of AI capabilities (Pause)
Reactive: Society keeps rough pace with AI (Muddle, Aligned)
Overwhelmed: Society can’t keep up (Catastrophe, parts of Multipolar)

Critical Branch Points

Certain decisions and events could significantly shift which scenario we end up in:

Near-Term (2024-2027)

Scaling continuation: Do we hit capability walls or keep scaling?
Catastrophic AI incident: Does a major AI accident occur?
Governance momentum: Do AI safety policies gain or lose steam?
China-US dynamics: Cooperation or intensifying competition?
Alignment breakthroughs: Do we make progress on core alignment problems?

Medium-Term (2027-2032)

AGI threshold: Do we develop systems that qualify as AGI?
Deployment decisions: How are powerful AI systems deployed?
International agreements: Can nations coordinate on AI development?
Economic impacts: How disruptive is AI to employment and growth?
Safety culture: Do leading labs maintain or abandon safety commitments?

Long-Term (2032-2040)

Superintelligence: Do we develop systems significantly beyond human intelligence?
Value lock-in: Do certain values or systems become entrenched?
Existential outcomes: Do we avoid catastrophic failures?

Using Scenario Analysis

❓Key Questions

Which scenario do you think is most likely? Why?

What evidence would make you update toward a different scenario?

Which interventions are valuable across multiple scenarios?

What early warning signs should we watch for?

How should we allocate resources given scenario uncertainty?

Intervention Robustness Matrix

Different interventions have different value depending on which scenario unfolds. This matrix helps identify robust interventions (valuable across scenarios) vs. scenario-specific bets.

Technical Interventions

Intervention	Aligned AGI	Slow Muddle	Catastrophe	Multipolar	Pause
Interpretability research	✅ Accelerates	✅ Ongoing value	⚠️ Maybe too late	✅ Useful	✅ Useful if resumed
RLHF/Constitutional AI	✅ Core technique	✅ Standard practice	❌ Didn’t scale	✅ Widely used	✅ Improved during pause
AI Control research	⚠️ Less needed	✅ Valuable	✅ Critical if possible	✅ Necessary	⚠️ Less urgent
Capability evals	✅ Enabled safe deployment	✅ Standard practice	⚠️ Didn’t prevent	✅ Arms race tool	✅ Monitoring tool
Agent foundations theory	⚠️ Not needed	⚠️ Uncertain value	❌ Too slow	⚠️ Uncertain	✅ Time to develop

Most robust technical intervention: Interpretability (valuable in 4/5 scenarios)

Governance Interventions

Intervention	Aligned AGI	Slow Muddle	Catastrophe	Multipolar	Pause
Compute governance	⚠️ Enabled control	✅ Key tool	⚠️ Maybe bypassed	✅ Critical	✅ Enables pause
International coordination	✅ Managed transition	✅ Prevented racing	❌ Failed	❌ Failed	✅ Made it possible
Lab safety standards	✅ Industry norm	✅ Helpful	⚠️ Insufficient	⚠️ Inconsistent	✅ Strengthened
Public advocacy	✅ Built support	✅ Maintained pressure	⚠️ Too late	⚠️ Polarized	✅ Created mandate
Liability frameworks	⚠️ Less needed	✅ Useful deterrent	❌ Irrelevant	✅ Some effect	✅ Incentive alignment

Most robust governance intervention: Compute governance (valuable in 4/5 scenarios)

Career/Personal Interventions

Intervention	Aligned AGI	Slow Muddle	Catastrophe	Multipolar	Pause
Safety research at lab	✅ Contributed	✅ Valuable	⚠️ Accelerated capabilities?	⚠️ Mixed	❌ Lab stopped
Independent safety research	⚠️ Less resourced	✅ Complementary	⚠️ Insufficient	✅ Independent voice	✅ Continued
Policy/governance work	✅ Shaped outcome	✅ Core contribution	⚠️ Didn’t prevent	✅ Critical	✅ Made it happen
Building alternative orgs	⚠️ Less needed	✅ Diversified field	⚠️ Too late	✅ Needed	✅ Essential
Public communication	✅ Built consensus	✅ Maintained awareness	⚠️ Warning ignored	⚠️ Weaponized	✅ Created mandate

Most robust career choice: Policy/governance work (valuable in 4/5 scenarios)

Reading the Matrix

Legend:

✅ = Clearly valuable in this scenario
⚠️ = Uncertain or mixed value
❌ = Not valuable or counterproductive

How to use this:

If you’re highly uncertain about scenarios: Choose robust interventions (many ✅)
If you have strong scenario beliefs: Weight by your probability estimates
If you want to hedge: Diversify across robust + scenario-specific interventions

Scenario-Weighted Expected Value

If you assign rough probabilities to scenarios, you can estimate intervention value:

Intervention	Example Weighting*	Robustness Score
Interpretability research	0.8	High
Compute governance	0.75	High
Policy/governance careers	0.75	High
International coordination	0.6	Medium
Safety research at frontier lab	0.5	Medium (controversial)
Agent foundations theory	0.4	Low

*Assuming equal scenario probabilities; your mileage will vary based on your beliefs.

Key Insight: Scenario Uncertainty Favors Governance

Technical interventions tend to be more scenario-specific (great if alignment works, useless if it doesn’t). Governance interventions tend to be more robust (useful for managing whatever happens).

If you’re uncertain which technical approach will work: Governance may be a better bet.

If you’re confident in a technical approach: Direct technical work may have higher upside.

Limitations of This Framework

What these scenarios don’t capture:

Black swan events: Truly unexpected developments we can’t anticipate
Technological surprises: Novel AI architectures or approaches
Non-AI factors: Climate change, pandemics, other existential risks
Detailed timelines: Specific years when events occur
Probability distributions: More nuanced likelihoods than rough ranges

Important caveats:

Scenarios are illustrative, not exhaustive - other futures are possible
Probabilities are subjective and should be debated, not taken as facts
Path dependence matters - early choices affect later possibilities
Scenarios can blend - we might see elements of multiple futures
The future is not fixed - our actions can shift probabilities