Parameter Interaction Network

Overview

AI safety parameters don't exist in isolation—they form a complex web of causal relationships where changes to one parameter ripple through the system. This model maps these interactions to identify leverage points, feedback loops, and critical dependencies.

Core insight: The parameter space clusters into four interconnected groups: (1) epistemic/trust parameters, (2) governance/coordination parameters, (3) technical safety parameters, and (4) exposure/threat parameters. Interventions on "hub" parameters like epistemic-health and institutional-quality propagate effects across multiple clusters.

Understanding these interactions matters for intervention design. Targeting isolated parameters yields limited returns; targeting hub parameters or breaking negative feedback loops offers higher leverage.

Conceptual Framework

Network Structure

The 22 parameters form a directed graph where edges represent causal influence. We distinguish three types of relationships:

flowchart TD
  subgraph Epistemic["Epistemic Cluster"]
      EH[Epistemic Health]
      IA[Information Authenticity]
      ST[Societal Trust]
      RC[Reality Coherence]
      PA[Preference Authenticity]
  end

  subgraph Governance["Governance Cluster"]
      IQ[Institutional Quality]
      RCap[Regulatory Capacity]
      IC[International Coordination]
      CC[Coordination Capacity]
      SCStr[Safety Culture Strength]
  end

  subgraph Technical["Technical Safety Cluster"]
      AR[Alignment Robustness]
      SCG[Safety-Capability Gap]
      RI[Racing Intensity]
      HO[Human Oversight Quality]
      IntC[Interpretability Coverage]
  end

  subgraph Exposure["Threat Exposure Cluster"]
      BTE[Biological Threat Exposure]
      CTE[Cyber Threat Exposure]
      ACC[AI Control Concentration]
      ES[Economic Stability]
      HA[Human Agency]
      SR[Societal Resilience]
      HE[Human Expertise]
  end

  EH --> ST
  IA --> EH
  ST --> IQ
  IQ --> RCap
  RI --> SCStr
  SCStr --> AR
  AR --> SCG
  HO --> AR
  ACC --> HA
  ES --> SR
  HE --> HO

Parameter Dependency Matrix

Core Causal Relationships

Influence Scores

Counting direct outgoing edges weighted by strength:

Feedback Loops

Identified Loops

The network contains 7 major feedback loops:

Loop Dynamics

flowchart TD
  subgraph ViciousLoop["Racing-Safety Spiral"]
      RI2[Racing Intensity ↑] --> |reduces| SCS[Safety Culture ↓]
      SCS --> |enables| RI2
  end

  subgraph VirtuousLoop["Trust-Institution Cycle"]
      ST2[Societal Trust ↑] --> |improves| IQ2[Institutional Quality ↑]
      IQ2 --> |enables| EH2[Epistemic Health ↑]
      EH2 --> |builds| ST2
  end

  subgraph DampeningLoop["Regulatory Response"]
      RI3[Racing Intensity ↑] --> |causes| ACC2[Accidents]
      ACC2 --> |triggers| RC2[Regulatory Capacity ↑]
      RC2 --> |reduces| RI3
  end

Racing-Safety Spiral: As racing intensifies, labs cut safety investments to maintain competitive position. Lower safety culture further normalizes speed-first decisions, intensifying the race. This loop operates on monthly timescales and is currently active in frontier AI development.

Trust-Institution Cycle: When societal trust is high, institutions attract talent and funding, improving their quality. Better institutions produce more reliable information, improving epistemic health, which feeds back to trust. This virtuous cycle takes years to establish but is self-reinforcing once started.

Expertise Erosion Loop: The most dangerous long-term loop. As humans defer to AI systems, expertise atrophies. Lower expertise reduces oversight quality, which eventually leads to alignment failures. Each failure damages the human knowledge base further. This loop operates over decades and may be effectively irreversible.

Cluster Analysis

Parameter Clusters

Cross-Cluster Dependencies

The clusters form a rough hierarchy:

This hierarchy suggests interventions should prioritize upstream clusters. Improving epistemic-health propagates through governance improvements to technical safety to reduced threat exposure. However, the time lags mean upstream interventions require patience—direct technical interventions may be necessary for near-term risk reduction.

Scenario Analysis

Cluster Degradation Scenarios

Positive Intervention Scenarios

Strategic Implications

High-Leverage Intervention Points

Based on network centrality and feedback loop positions:

Intervention Timing

Different parameters have different optimal intervention windows:

Limitations

1. Causal uncertainty: Many relationships are theorized rather than empirically confirmed. The strength estimates are order-of-magnitude guesses.

2. Missing parameters: The 22 parameters don't capture everything relevant. Military dynamics, public opinion volatility, and AI capability trajectories are underrepresented.

3. Static structure: The network structure itself may change as AI capabilities advance. New feedback loops may emerge.

4. Aggregate treatment: Each parameter aggregates many underlying variables. "Institutional quality" obscures differences between regulatory agencies, courts, and legislatures.

5. Linear approximation: Relationships may be non-linear with threshold effects not captured by simple positive/negative coding.

Related Models

• Risk Interaction NetworkAnalysisAI Risk Interaction Network ModelSystematic analysis identifying racing dynamics as a hub risk enabling 8 downstream risks with 2-5x amplification, and showing compound risk scenarios create 3-8x higher catastrophic probabilities ...Quality: 64/100 - Similar network approach for risks
• Risk Cascade PathwaysAnalysisAI Risk Cascade Pathways ModelIdentifies 5 AI risk cascade pathways with probabilities of 1-45% for catastrophic outcomes over 5-50 year timelines, finding racing dynamics as the highest leverage intervention point (80-90% trig...Quality: 67/100 - How risks propagate
• Epistemic Collapse ThresholdAnalysisEpistemic Collapse Threshold ModelModels epistemic collapse as threshold phenomenon where society loses ability to establish shared facts, estimating 75-80% combined probability of collapse via authentication failure (35-45%), pola...Quality: 50/100 - Deep dive on epistemic parameters
• Trust Cascade ModelAnalysisTrust Cascade Failure ModelModels institutional trust as a network contagion problem, finding cascades become irreversible below 30-40% trust thresholds and that AI multiplies attack effectiveness 60-5000x while degrading de...Quality: 58/100 - Trust dynamics in detail