Part III · Chapter 6

Power, Legitimacy and the Boundaries of AI

Political power cannot be reduced to the final actor who signs or votes. Decision, agenda, procedural, definitional, structural, epistemic, and preference-shaping forms of power illuminate different mechanisms.

How Political Science Learned to KnowDecision Dossier VILiving companion

Learning objectives

What this chapter asks you to be able to do.

Distinguish direct decision power from agenda, procedural, definitional, structural, and preference-shaping forms of power.
Compare Dahl, Bachrach and Baratz, Lukes, and later relational accounts of power without treating any one definition as uncontested.
Distinguish influence, persuasion, expertise, coercion, domination, legitimate authority, and institutional dependence.
Explain sociological and normative meanings of legitimacy and distinguish legitimacy from legality, popularity, performance, and epistemic credibility.
Evaluate when deference to expertise is reasonable and when expert dependence can become political insulation.
Identify how an AI system can acquire consequential de facto influence without possessing legal or sovereign authority.
Explain why a nominal "human in the loop" may be insufficient for meaningful human authorization.
Separate causal contribution, professional responsibility, public accountability, legal liability, and constitutional authorization in AI-assisted decision processes.
Evaluate the current NousPolis separation-of-powers, human-governance, provenance, independence, review, and maturity rules as proposed safeguards rather than accomplished facts.
Extend the Asterbridge Decision Dossier with a Power, Legitimacy and AI Role Record.

Chapter summary

The argument in inspectable form.

Dahl's relational account makes power empirically tractable through observable effects on choices but can miss conflicts prevented from becoming visible.

Bachrach and Baratz identify a second face of power in agenda control and nondecision-making, where institutions restrict which grievances and alternatives become subjects of decision.

Lukes's third dimension asks how power can shape perceptions and preferences, while raising a serious problem about who may define another person's "real interests."

Structural and relational approaches shift attention from a single powerful actor toward social boundaries, institutional relationships, dependencies, and capacities.

Influence, expertise, persuasion, coercion, domination, de facto power, and legitimate authority are not synonyms. Political analysis should identify the mechanism before judging it.

AI can exercise consequential influence through search, classification, framing, routing, summarization, scoring, option generation, recommendation, stopping, publication, and reopening even when it possesses no legal authority.

Capability, influence, procedural power, and formal authority must be recorded separately. A powerful recommender need not be a sovereign, and a formally authorized institution need not be the most influential actor in practice.

Legitimacy has sociological and normative meanings and should be distinguished from legality, public acceptance, performance, procedural validity, democratic authorization, and epistemic credibility.

Expertise is indispensable to modern governance, but epistemic deference does not automatically create political authority. Technocratic risks arise from institutional insulation, not from the mere existence of specialists.

A "human in the loop" is not automatically meaningful human control. Override quality depends on authority, evidence access, time, competence, alternatives, incentives, dissent visibility, and review.

Automation bias is a documented risk in some settings, but public-sector evidence does not justify assuming universal algorithmic deference. Oversight design should be empirically tested.

Responsibility should be separated into causal contribution, professional responsibility, legal liability, political accountability, and constitutional authorization.

Contestability requires reasons, uncertainty disclosure, visible dissent, provenance, alternative framings, explicit authority status, review routes, meaningful override, and reopening conditions proportionate to consequence.

Current NousPolis architecture directly recognizes procedural and definitional power, but N0 separation and governance documents are not evidence that the controls are executable or institutionally independent.

Decision Dossier VI records who shapes the conditions under which authorized humans choose, including de jure/de facto gaps, dependencies, legitimacy claims, override quality, dissent visibility, accountability paths, and reopening triggers.

Key terms

Concepts to carry forward.

power: The capacity or social relation through which actors, institutions, rules, or structures shape what others do, what choices are available, what issues are considered, or how political subjects and interests are constituted. Competing theories disagree about which of these effects should count as power.
influence: An effect on another actor's beliefs, attention, preferences, or conduct. Influence can arise through persuasion, expertise, example, incentives, dependence, manipulation, or coercion and is not automatically illegitimate.
authority: A recognized or justified right to make, apply, or enforce decisions within a domain. Authority is not identical to de facto influence or expertise.
decision power: Power exercised in observable conflict over an issue, where one actor affects the choice or conduct of another.
agenda power: Power over which issues, alternatives, claims, or conflicts receive decision-making attention and which remain excluded, deferred, or treated as nonissues.
procedural power: Power exercised through sequencing, routing, evidence scope, panel composition, thresholds, stopping rules, aggregation, review, publication, or other rules governing how a decision is produced.
definitional power: Power exercised through classifications, labels, categories, metrics, and thresholds that determine what an object is and which safeguards or institutional pathways apply.
structural power: Power arising from durable social and institutional relations that shape actors' positions, dependencies, capacities, and feasible choices, including effects not traceable to one deliberate command.
epistemic power: The capacity to shape what is treated as credible knowledge, relevant evidence, authoritative interpretation, or a reasonable basis for action.
legitimacy: A contested concept concerning the rightfulness, acceptability, or justification of rule or institutions. Sociological legitimacy concerns beliefs in rightfulness; normative legitimacy concerns whether power is actually justified under defensible standards.
technocracy: A form or tendency of governance in which specialized technical expertise acquires substantial authority or insulation in public decision-making. The term can describe different institutional arrangements and should not be used as a synonym for expertise itself.
contestability: The practical ability to question, challenge, obtain reasons for, review, and potentially alter a decision, classification, procedure, or recommendation.
meaningful override: A human capacity to depart from an automated or expert recommendation that is real rather than ceremonial: the human has authority, information, time, competence, alternatives, and institutional permission to reject or revise the recommendation.
automation bias: A tendency, demonstrated in some human-automation settings, to rely excessively on automated advice or to miss contradictory information. Its prevalence and strength vary by context and should not be assumed in every public-sector use of AI.
accountability: A relationship in which an actor is obliged to explain and justify conduct to a forum able to question, judge, and potentially impose consequences; broader uses of the term also include political, legal, professional, and institutional forms of answerability.
de facto power: Consequential influence exercised in practice, whether or not the actor possesses a formal legal power to produce the outcome.
de jure authority: Authority recognized by the applicable legal or constitutional order.

Review & discussion

Questions for seminar, revision, or assessment.

Why is a map of formal authority insufficient for understanding political power in the Asterbridge briefing case?
What evidence would a Dahlian analysis require before concluding that the AI system changed a councillor's decision?
Give one example of a nondecision that does not require any actor to intentionally suppress an issue.
What is gained by Lukes's third dimension of power, and what paternalistic danger does it create for the researcher?
How does structural power differ from a direct command? Give an example involving institutional dependence on a technical system.
When does a classification become a political act rather than merely an administrative convenience?
Construct a case in which an AI system has high de facto influence but zero formal authority. Then construct the reverse.
Why does a high rate of agreement between humans and an AI recommender fail to prove automation bias?
Distinguish sociological legitimacy from normative legitimacy using one government institution or fictional Asterbridge body.
Can a process be legally valid but democratically contested? Can it be publicly popular but normatively unjustified? Explain.
What is the difference between epistemic credibility and political legitimacy?
When is deference to expertise rational? What institutional conditions turn deference into problematic insulation?
Why is "human in the loop" a weaker claim than "meaningful human authorization"?
What evidence would convince you that an override mechanism is real rather than ceremonial?
Compare the automation-bias findings from aviation or decision-support research with Alon-Barkat and Busuioc's public-sector experiments. Why should institutional design not assume one universal human response to algorithms?
In an AI-assisted policy failure, why might the causal contributor, legally liable party, politically accountable institution, and constitutional authorizer be different actors?
Design a twelve-page summary rule that preserves meaningful dissent without making the brief unreadable. What political choices remain in your design?
Does provenance reduce power, expose power, or both? Explain why a perfectly reproducible process can still be illegitimate.
How could several role-separated AI agents remain controlled by one underlying principal? What independence claim would be justified?
Can an orchestrator that never casts a substantive vote become the most powerful actor in a decision process? State the conditions under which your answer would be yes and the evidence needed to demonstrate it.
Under what conditions could repeated human approvals amount to legitimacy laundering? Under what conditions would the same approval pattern be unobjectionable?
Suppose an extremely accurate AI recommendation conflicts with a procedurally valid human decision based on rights or distribution. What exactly is the disagreement about?
Which safeguard in current NousPolis architecture most directly addresses agenda or procedural power? What is its strongest remaining weakness at N0?
Identify one form of political power that Decision Dossier VI could still fail to capture. How would you amend the dossier?

Continue into the literature.

For the modern debate over power, begin with Dahl's compact relational account (1957), then read Bachrach and Baratz (1962) and Lukes's second edition of Power: A Radical View (2005) as successive expansions and critiques of what observable decision-making can reveal. Hayward (2000) is especially useful for moving away from a search for a single "face" of power, while Barnett and Duvall (2005) provide a clear taxonomy of compulsory, institutional, structural, and productive power in international politics. For legitimacy, Weber (1978) remains foundational for sociological analysis, while Buchanan (2002) makes the normative problem explicit. Scharpf (1999) and Schmidt (2013) are valuable for thinking about input, output, and throughput legitimacy without treating those categories as a complete theory. On expertise and democracy, pair Fischer's critique of technocracy (1990) with Estlund (2008) and Schwartzberg (2015). For human oversight of automated systems, read Green (2022) alongside Parasuraman and Manzey (2010) and Alon-Barkat and Busuioc (2023) so that institutional warnings are read together with evidence about context-dependent human behavior. Bovens (2007) provides a disciplined framework for accountability, and Santoni de Sio and Mecacci (2021) show why AI-related responsibility gaps should be disaggregated rather than treated as one problem.

Companion, not replacement

The textbook carries the complete argument.

This page reproduces the chapter's study and navigation layer from the current living manuscript. The Asterbridge narrative, historical and methodological argument, figures, Political Science Lens boxes, NousPolis Canon crosswalks, labs, red-team exercises, and N5 Practice boundaries remain in the canonical textbook rather than being republished wholesale here.