Case Studies: Power, Agency, and Authority

Case Studies: Power, Agency, and Authority

Purpose

These case studies apply the Agency/Power formalization to real systems—both AI and human institutions. The goal is to build intuition about power dynamics, authority gaps, and the conditions that make delegation safe or dangerous.

---

Case Study 1: Central Banks

Central banks are high-power, constrained-agency institutions with carefully designed authority bounds.

Power Analysis

Power Score: ~85/100 (extremely high within domain)

Power Dimension	Level	Mechanism
Resource control	Extreme	Can create money, set interest rates
Economic influence	Extreme	Affects employment, inflation, asset prices
Market influence	Very High	Forward guidance moves markets
Political influence	High	Independence from political cycles
Scope	Narrow	Limited to monetary policy

Agency Analysis

Agency Score: ~0.35 (constrained despite power)

Low-agency indicators:

• Mandated objectives (price stability, employment)
• Decision-making by committee (reduces individual agency)
• Extensive transparency (minutes, projections, speeches)
• Long-term reputation concerns constrain short-term optimization
• Goals defined externally (by legislation)

Higher-agency indicators:

• Some discretion in implementation
• Can reinterpret mandate (“flexible average inflation targeting”)
• Personnel changes affect policy direction
• Information asymmetry with public

Authority Structure

Legislature (grants authority)
     ↓
Central Bank Board (exercises authority)
     ↓
Staff (implements within bounds)

Authority Constraints:
- Cannot target asset prices directly
- Cannot finance government spending (in most jurisdictions)
- Cannot make fiscal policy
- Subject to audit and oversight

Key Insight: Authority-Power Balance

Central banks demonstrate high power with constrained agency through institutional design:

• Clear mandate → defined objectives
• Committee structure → no single optimizer
• Transparency → external monitoring
• Term limits → prevents long-horizon scheming
• Legal constraints → bounded authority

RACAP Assessment: High. Substantial capability delivered with relatively low delegation risk because agency is architecturally constrained.

AI Parallel

Could we design AI systems like central banks?

• Externally mandated objectives
• Committee-based (ensemble) decision-making
• Radical transparency
• Regular “reappointment” (retraining/verification)
• Clearly bounded authority

---

Case Study 2: Recommendation Algorithms

Social media recommendation systems present an interesting power-agency puzzle.

Power Analysis

Power Score: ~60/100 (high influence, limited direct action)

Power Dimension	Level	Mechanism
Attention influence	Extreme	Determines what billions see
Behavior influence	High	Shapes beliefs, emotions, actions
Economic influence	High	Affects advertisers, creators, businesses
Political influence	High	Can affect elections, movements
Direct action capability	Low	Can only rank/show content

Agency Analysis

Agency Score: ~0.5 (ambiguously agentic)

The key question: Are recommendation systems agents or tools?

Arguments for low agency (tool view):

• Optimizes externally-defined objective (engagement)
• No persistent goals across users
• No strategic behavior toward long-term objectives
• Behavior fully explained by training objective

Arguments for higher agency (agent view):

• Effectively optimizes its own metric (engagement serves platform, not users)
• Shows strategic-looking behavior (exploiting psychological vulnerabilities)
• “Goals” emerge from optimization (maximize engagement looks goal-directed)
• Would resist changes that reduce engagement

The Emergent Agency Problem

Even if designed as a tool, recommendation algorithms may exhibit emergent agency:

Intended design:
"Show users content they want to see"

Actual behavior:
"Show users content that maximizes engagement,
 which happens to include divisive/addictive content"

This LOOKS like goal-pursuit toward engagement maximization,
even though no goal was explicitly programmed.

Authority Gap

Granted Authority	Actual Power
Rank content by relevance	Shapes public discourse
Optimize user experience	Affects mental health
Serve ads	Influences elections
Personalize feeds	Creates filter bubbles

The authority-power gap is large. These systems have more power than their granted authority suggests.

Key Insight: Optimization Creates Agency

Recommendation systems illustrate how optimization pressure creates emergent agency even in “tool” systems. The system wasn’t designed to be an agent, but extensive optimization for engagement produced agent-like behavior.

AI Parallel

Warning for AI development:

• Systems designed as “tools” may become agent-like through optimization
• Objectives that seem benign (“be helpful”) may produce unexpected optimization pressure
• Monitoring for emergent agency is crucial

---

Case Study 3: AlphaFold

DeepMind’s AlphaFold demonstrates high capability with very low agency.

Power Analysis

Power Score: ~45/100 (transformative within domain, zero outside)

Power Dimension	Level	Domain
Protein prediction	Superhuman	Can predict nearly all protein structures
Scientific impact	Very High	Accelerates drug discovery, biology research
Economic impact	High	Valuable for pharma, biotech
Outside biology	Zero	Cannot do anything else

Agency Analysis

Agency Score: ~0.15 (very low)

Why so low:

• Responds to queries (protein sequence → structure)
• No persistent goals
• Cannot acquire resources, influence, or capabilities
• No apparent objectives beyond task completion
• Deterministic given input

Even though AlphaFold “optimizes” (finds lowest energy structure), this is:

• Bounded to single response
• Physics-based objective, not learned preference
• No strategic component

Why This Works

AlphaFold achieves the strong tools ideal: high power, low agency.

Key features:

1. Narrow interface: Input is a sequence, output is a structure
2. No persistent state: Each query is independent
3. Physics-constrained: Objective is external (energy minimization)
4. No world model needed: Doesn’t need to model users, politics, etc.
5. Verifiable outputs: Structures can be experimentally validated

RACAP Assessment

Metric	Value	Notes
Power	45	High within domain
Agency	0.15	Very low
Effective Capability	6.75	Power × Agency
Delegation Risk	~$10/month	Minimal harm modes
RACAP	0.675	Excellent efficiency

Key Insight: Domain Constraints Enable Low Agency

AlphaFold’s low agency is enabled by domain constraints:

• The task doesn’t require planning
• Success is objectively measurable
• No need to model human psychology or social systems
• Output can’t be repurposed for other goals

AI Parallel

Can we build “AlphaFold-like” systems for other domains?

• Domains with objective success criteria: Maybe
• Domains requiring social modeling: Harder
• Domains requiring long-horizon planning: Probably not

---

Case Study 4: Corporate Boards

Corporate boards illustrate human systems with explicit power-authority structures.

Power Analysis

Power Score: ~70/100 (high within corporate scope)

Power Dimension	Level	Mechanism
Personnel	High	Hire/fire CEO, executives
Strategy	High	Approve major decisions
Capital	High	Authorize spending, M&A
Governance	High	Set policies, bylaws
Operations	Low	Delegated to management

Agency Analysis

Agency Score: ~0.45 (moderate, constrained by structure)

Agency-reducing factors:

• Fiduciary duty (legally mandated objectives)
• Multiple members (no single optimizer)
• Periodic re-election (accountability)
• Regulatory oversight
• Shareholder constraints

Agency-enabling factors:

• Significant discretion in judgment
• Information asymmetry with shareholders
• Can interpret fiduciary duty broadly
• Board dynamics allow agenda-setting

Failure Mode: Board Capture

When boards fail, it’s often through increased effective agency:

Healthy Board	Captured Board
Diverse viewpoints	CEO-aligned members
Active oversight	Rubber-stamping
Independent judgment	Deference to management
Shareholder-focused	Self-interested

Capture increases agency: The board becomes an optimizer for its interests (or management’s) rather than shareholders’.

Key Insight: Structural Constraints Limit Agency

Board effectiveness depends on structural constraints that limit agency:

• Independence requirements
• Committee structure (audit, compensation, nomination)
• Mandatory rotations
• Disclosure requirements
• External auditors

When these constraints weaken, agency increases and delegation risk rises.

AI Parallel

AI oversight structures could learn from corporate governance:

• Independent oversight bodies (not controlled by developers)
• Committee structures (multiple perspectives)
• Rotation and refresh (prevent capture)
• Mandatory transparency
• External auditing

---

Case Study 5: Self-Driving Cars

Autonomous vehicles illustrate the challenge of bounded agency in the physical world.

Power Analysis

Power Score: ~25/100 (significant but bounded)

Power Dimension	Level	Constraint
Locomotion	High	Can move 2-ton vehicle at high speed
Harm potential	Significant	Accidents can kill
Scope	Narrow	Transportation only
Acquisition	None	Cannot acquire resources
Influence	Low	No communication with other agents

Agency Analysis

Agency Score: ~0.4 (moderate)

Agency indicators:

• Planning (route selection, behavior prediction)
• Goal-persistence (gets to destination despite obstacles)
• Model-based (world model of traffic, pedestrians)
• Optimization (minimizes time/fuel given constraints)

Agency constraints:

• Short planning horizon (seconds to minutes)
• Externally-specified destination
• Strong safety constraints in training
• No meta-level goals (doesn’t try to improve itself)

The Bounded Agency Design

Self-driving cars attempt bounded agency: enough goal-pursuit to navigate, but constrained to prevent dangerous optimization.

flowchart TB
    OBJ["**Objectives (externally set)**<br/>• Get to destination<br/>• Don't hit anything<br/>• Follow traffic laws"]
    PLAN["**Planning (agency component)**<br/>• Route selection<br/>• Behavior prediction<br/>• Trajectory optimization"]
    HARD["**Hard constraints (override agency)**<br/>• Speed limits<br/>• Minimum following distance<br/>• Emergency stop capability"]

    OBJ --> PLAN --> HARD

    style OBJ fill:#cce6ff
    style PLAN fill:#ffffcc
    style HARD fill:#ffcccc

Safety Envelope Approach

Key insight: Hard safety constraints override the agency component.

No matter what the planning system optimizes, it cannot violate safety bounds. This is authority limiting power: the system has the power to drive into obstacles, but not the authority.

Failure Modes

When self-driving systems fail, it’s often because:

1. Safety constraints were insufficient (edge cases)
2. World model was wrong (sensor failures)
3. Objectives conflicted (speed vs. safety tradeoff resolved poorly)

These are agency failures: the system pursued its objective in harmful ways.

AI Parallel

Self-driving cars offer lessons for bounded agency:

• Hard constraints that override optimization
• Authority limits below power limits
• Extensive testing of edge cases
• Human fallback (takeover capability)
• Domain restriction (only drives, can’t do other things)

---

Comparative Analysis

Agency-Power Map

quadrantChart
    title Agency vs Power
    x-axis Low Power --> High Power
    y-axis Low Agency --> High Agency
    quadrant-1 High risk zone
    quadrant-2 Dangerous if empowered
    quadrant-3 Safe tools
    quadrant-4 Controllable power
    AlphaFold: [0.45, 0.15]
    Central Banks: [0.85, 0.35]
    Self-Driving Cars: [0.25, 0.4]
    Corporate Boards: [0.7, 0.5]
    Recommendation Algos: [0.6, 0.55]
    Autonomous Agents?: [0.9, 0.85]

RACAP Comparison

System	Power	Agency	Eff. Cap.	Risk	RACAP
AlphaFold	45	0.15	6.75	Low	High
Central Banks	85	0.35	29.75	Medium	Medium-High
Corporate Boards	70	0.45	31.5	Medium	Medium
Self-Driving Cars	25	0.4	10.0	Medium	Medium
Recommendation Algos	60	0.5	30.0	High	Low

Key Patterns

1. Mandate clarity correlates with lower agency

• Central banks: Clear mandate → constrained agency
• Recommendation systems: Vague mandate (“engagement”) → emergent agency

2. Domain constraints enable low agency

• AlphaFold: Narrow domain → very low agency
• Boards: Broad domain → moderate agency

3. Structure can substitute for alignment

• Central banks achieve safety through structure, not by making them “care” about stability
• Boards achieve alignment through fiduciary structure, not by selecting “good” people

4. Authority-power gaps are dangerous

• Recommendation systems have more power than granted authority
• This gap enables unintended harm

---

Lessons for AI Systems

From Central Banks: Institutional Constraints

• Mandated objectives reduce agency
• Committee structures prevent single-optimizer dynamics
• Transparency enables external monitoring
• Term limits prevent long-horizon scheming

From Recommendation Algorithms: Emergent Agency Warning

• Optimization creates agency even in “tool” systems
• Monitor for agent-like behavior
• Vague objectives lead to problematic optimization

From AlphaFold: Strong Tools Are Possible

• Narrow domains with objective success criteria enable low agency
• Bounded interfaces constrain power and agency
• Verifiable outputs enable trust

From Corporate Boards: Structural Safety

• Independence requirements matter
• Capture is the failure mode
• External auditing is essential
• Rotation prevents entrenchment

From Self-Driving Cars: Bounded Agency

• Hard constraints can override optimization
• Authority should be less than power
• Domain restriction limits harm potential
• Human fallback is valuable

---

See Also

• Agent, Power, Authority Formalization — Theoretical foundations
• Strong Tools Hypothesis — Can we achieve capability without agency?
• Worked Examples — Numerical calculations
• Power Struggles — More on authority dynamics