Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Chapter 17: Sensitivity Analysis

Which Variables Matter Most: Understanding Parameter Influence

Not all uncertainties are created equal. Some parameters dramatically affect outcomes while others barely register. This chapter identifies which variables drive our results and where small changes create large differences in the future.

Methodology

Global Sensitivity Analysis

We systematically vary each parameter while holding others constant, measuring:

  • First-order effects: Direct parameter impact
  • Interaction effects: Combined parameter impacts
  • Total sensitivity: All effects combined
  • Threshold effects: Non-linear transitions

Sobol Indices

Using Sobol sensitivity analysis, we decompose variance:

  • Si: First-order sensitivity index
  • STi: Total-order sensitivity index
  • STi - Si: Higher-order interaction effects

Parameter Rankings

Highest Sensitivity Parameters (STi > 0.20)

1. H1 Prior Probability (AI Progress)

  • First-order: 0.187
  • Total: 0.241
  • Interactions: 0.054
  • Impact: 24.1% of total variance

Interpretation: AI progress rate is the single most influential factor. Small changes in this probability dramatically affect all scenarios.

2. H5 Prior Probability (Development Model)

  • First-order: 0.134
  • Total: 0.218
  • Interactions: 0.084
  • Impact: 21.8% of total variance

Interpretation: Whether AI development centralizes is nearly as important as progress rate itself. High interaction effects suggest it amplifies other factors.

3. Causal Strength Multiplier

  • First-order: 0.098
  • Total: 0.201
  • Interactions: 0.103
  • Impact: 20.1% of total variance

Interpretation: How strongly hypotheses influence each other matters enormously. Conservative vs aggressive models yield very different futures.

High Sensitivity Parameters (STi 0.10-0.20)

4. H6 Prior Probability (Governance)

  • Total: 0.156
  • Impact: Democratic vs authoritarian outcomes significantly affect scenario probabilities

5. H3 Prior Probability (Employment)

  • Total: 0.143
  • Impact: Complement vs displacement effects ripple throughout analysis

6. H2-H3 Causal Strength (AGI → Employment)

  • Total: 0.128
  • Impact: How AGI affects work is crucial for determining societal response

Medium Sensitivity Parameters (STi 0.05-0.10)

7. H2 Prior Probability (AGI Achievement): 0.098 8. H5-H6 Causal Strength (Centralization → Authoritarianism): 0.089 9. Temporal Discount Rate: 0.078 10. H4 Prior Probability (Safety): 0.067

Low Sensitivity Parameters (STi < 0.05)

11-20. Various individual causal relationships: 0.015-0.045 21-25. Evidence quality adjustments: 0.008-0.025

Threshold Analysis

Critical Thresholds

H1 (AI Progress) Threshold: 75%

  • Below 75%: Constrained Evolution becomes most likely
  • Above 75%: Adaptive Integration or Fragmented Disruption dominate
  • At 91.1% (our estimate): Strong acceleration likely

H5 (Centralization) Threshold: 60%

  • Below 60%: Distributed development scenarios viable
  • Above 60%: Centralization scenarios dominate
  • At 77.9% (our estimate): Extreme centralization likely

Causal Strength Threshold: 1.5x

  • Below 1.0x: Multiple scenarios remain viable
  • 1.0x-1.5x: Three-future structure emerges
  • Above 1.5x: Winner-take-all dynamics dominate

Interaction Effects

Strongest Interactions

H1 × H5 (Progress × Centralization)

  • Interaction strength: 0.067
  • Effect: Rapid progress drives centralization more than linear combination suggests
  • Mechanism: Compute requirements create winner-take-all dynamics

H5 × H6 (Centralization × Governance)

  • Interaction strength: 0.061
  • Effect: Centralization enables authoritarianism beyond simple correlation
  • Mechanism: Power concentration creates self-reinforcing dynamics

H2 × H3 (AGI × Employment)

  • Interaction strength: 0.044
  • Effect: AGI achievement makes displacement more likely than expected
  • Mechanism: General intelligence threatens broader job categories

Suppression Effects

Some parameters show negative interactions:

H4 × H6 (Safety × Governance)

  • Interaction: -0.021
  • Effect: Safety measures reduce authoritarian probability more than linear
  • Mechanism: Trust in AI reduces democracy-threatening crises

Scenario-Specific Sensitivity

Most Sensitive Scenarios

ABBABB (Rank 1, 11.59%)

  • Most sensitive to: H1, H5
  • Least sensitive to: H2, H4
  • Why: Core assumptions align perfectly with this scenario

AABABB (Rank 2, 9.21%)

  • Most sensitive to: H1, H2, H5
  • Moderately sensitive to all parameters
  • Why: Balanced across multiple dimensions

Least Sensitive Scenarios

Extreme scenarios (AAAAAA, BBBBBB):

  • Low sensitivity to all parameters
  • Already have very low probabilities
  • Small changes don’t affect much

Mixed scenarios with contradictory patterns:

  • Moderate sensitivity
  • Changes can shift them between clusters

Geographic Sensitivity

Parameter Importance by Region

Western Democracies:

  • Most sensitive to: H6 (governance), H4 (safety)
  • Least sensitive to: H1 (progress inevitable anyway)

Authoritarian Countries:

  • Most sensitive to: H1 (progress), H5 (centralization)
  • Least sensitive to: H6 (governance already decided)

Developing Nations:

  • Most sensitive to: H3 (employment), H5 (development model)
  • Medium sensitivity to others

Temporal Sensitivity Evolution

Sensitivity Changes Over Time

2025-2028: Maximum Sensitivity

Parameter       Early   Late    Change
H1 (Progress)   0.24    0.08    -67%
H5 (Central)    0.22    0.12    -45%
H6 (Govern)     0.16    0.21    +31%

Interpretation:

  • Technology parameters matter most early
  • Governance parameters matter most later
  • Economic parameters peak in middle period

Intervention Timing Implications

High-leverage early interventions:

  • Affecting AI progress trajectory
  • Influencing development model
  • Setting causal interaction patterns

High-leverage later interventions:

  • Democratic institution strengthening
  • Employment transition support
  • International governance coordination

Model Uncertainty

Sensitivity to Model Structure

Conservative vs Aggressive Models:

  • Conservative: More uniform probability distribution
  • Aggressive: More extreme scenarios dominate
  • Difference: Up to 40% probability shifts

Linear vs Nonlinear Causal Functions:

  • Linear: Gradual probability changes
  • Nonlinear: Threshold effects and jumps
  • Difference: Timing of transitions varies ±3 years

Practical Implications

For Decision Makers

Focus Attention On:

  1. AI progress trajectory (highest impact)
  2. Development centralization (high impact, controllable)
  3. Causal interaction strength (model assumptions matter)

Monitor Closely:

  1. Democratic institution health
  2. Employment displacement patterns
  3. AGI emergence indicators

Less Critical:

  • Individual safety incidents
  • Specific technical developments
  • Minor regulatory changes

For Researchers

Research Priorities:

  1. Better estimates of H1 and H5 probabilities
  2. Causal strength magnitudes
  3. Interaction mechanisms
  4. Threshold identification

Model Improvements:

  1. Dynamic causal strengths
  2. Geographic variation
  3. Feedback loop modeling
  4. Nonlinear relationships

For Activists

Campaign Priorities:

  1. Influence development model (H5)
  2. Strengthen democracy (H6)
  3. Shape progress narrative (H1)
  4. Moderate causal extremes

Less Effective:

  • Fighting individual technologies
  • Single-issue campaigns
  • Purely reactive responses

Robustness Tests

Leave-One-Out Analysis

Removing highest-sensitivity parameters:

  • Without H1: Three futures still emerge but probabilities shift dramatically
  • Without H5: Distribution becomes more uniform
  • Without causal interactions: Scenarios become independent

Conclusion: Core structure is robust but magnitudes are sensitive.

Alternative Parameterizations

Testing different prior distributions:

  • Uniform priors: Results similar but less extreme
  • Expert survey priors: Generally confirm our results
  • Historical analogy priors: Shift toward conservative scenarios

The Sensitivity Message

Key Insights

  1. AI progress rate dominates everything - Get this wrong and everything else is wrong
  2. Centralization is nearly as important - And more controllable
  3. Interactions matter enormously - Linear thinking underestimates effects
  4. Early parameters matter most - But timing switches to governance
  5. Most parameters don’t matter much - Focus on the few that do

Strategic Implications

For Maximum Impact:

  • Focus on highest-sensitivity parameters
  • Act when their sensitivity is highest
  • Understand interaction effects
  • Don’t waste time on low-impact variables

For Risk Management:

  • Monitor threshold approaches
  • Prepare for interaction effects
  • Build resilience to sensitivity changes
  • Maintain options as sensitivity shifts

The sensitivity analysis reveals that while the future is complex, influence is not equally distributed. A few key parameters drive most of the variation. Master these, and you master the future’s trajectory.

Next: Convergence Patterns →
Previous: Temporal Evolution ←