A 'Right-to-Act' protocol would place a deterministic veto at the point of AI execution, stopping any action if required conditions are unmet rather than letting high-confidence signals override failures; the approach promises safer reversibility but remains untested at scale and may impose practical trade-offs.

Main Finding

The paper identifies a distinct architectural gap — the Pre-Action Legitimacy Gap — and proposes a formal Right-to-Act primitive: a deterministic, non‑compensatory decision boundary that treats execution legitimacy as a set of required boolean constraints rather than a score. The author proves compensatory scoring systems (weighted scores with thresholds) cannot, in general, guarantee this form of per-decision legitimacy. A short case study (AI-driven account suspension) illustrates how a high-confidence, policy‑compliant decision can still be blocked under a Right-to-Act rule when a required structural condition is missing.

Key Points

• Problem statement
  • Current stacks (authorization, safety filters, governance, statistical certification) do not answer: “Should this AI-generated decision be allowed to exist as an executable event?”
  • This missing check is named the Pre-Action Legitimacy Gap.
• Formalization
  • Decision a is an abstract structured object (operation, context, target, scope, timing).
  • A finite set of required constraints C1...Cn are deterministic predicates Ci(a) ∈ {0,1}.
  • Right-to-Act decision function D(a) = ALLOW iff ∀i Ci(a)=1; otherwise D(a)=NON-ACTION (e.g., DEFER, ESCALATE, REQUEST-INFO).
  • Legitimacy is modeled as membership in feasible region F = {a : ∀i Ci(a)=1}.
• Non-compensatory property
  • A failed required constraint cannot be offset by satisfied constraints; any single failed Ci blocks execution.
  • Lemma: increasing other satisfied constraints without fixing a failed one does not change the NON-ACTION outcome.
• Non-equivalence theorem
  • Shows existence of decisions where compensatory scoring systems (S(a)=Σwixi(a) with threshold θ) allow execution while Right-to-Act rejects it.
  • Corollary: No choice of weights can turn a compensatory system into a true non-compensatory Right-to-Act system without introducing explicit hard constraints.
• Architectural placement and method
  • Right-to-Act is placed as a distinct boundary between AI-generated decisions and downstream validation/execution.
  • Proposed method: (1) enumerate decision classes, (2) define structured decision objects, (3) define required constraints, (4) define non-action outcomes, (5) evaluate via case studies.
• Case study: account suspension
  • Baseline: high aggregate score → immediate disable.
  • Right-to-Act: missing contextual verification constraint (Ccontext=0) → NON-ACTION (escalate or request info).
  • Demonstrates practical divergence and operational impact.
• Limits acknowledged
  • Paper is architectural/formal; it does not prescribe constraint content, calibration procedures, or implementation details.
  • Recognizes recursive-data/certification limits when training/eval data are AI-influenced.

Data & Methods

• Data
  • No empirical dataset is provided or used. The contribution is theoretical and architectural.
  • Uses a single illustrative case study (qualitative) rather than quantitative experimental data.
• Methods / formal apparatus
  • Minimal formal model: abstract action space A, structured action tuples, boolean predicates for constraints.
  • Right-to-Act decision rule expressed as a conjunctive boolean feasibility condition (hard constraints).
  • Definition of compensatory scoring systems as linear combinations of bounded features with threshold rule.
  • Theoretical proof (constructive counterexample) establishing non-equivalence between scoring systems and non-compensatory constraints.
  • Conceptual method for producing publishable Right-to-Act analyses (identification, constraint definition, mapping to non-action outcomes, case study comparison).
• Empirical testability notes (from paper)
  • The framework is designed to be compatible with heterogeneous AI components and testable via comparative case studies (compensatory baseline vs Right-to-Act boundary), but the paper does not implement such experiments.
• Methodological caveats
  • Constraint selection, representation, and enforcement costs are left as domain work.
  • Boolean reduction of complex judgments may require rich domain modeling and is not trivial.

Implications for AI Economics

• Frictions vs. efficiency trade-off
  • Introducing non-compensatory Right-to-Act constraints creates deliberate execution frictions (deferral, escalation), which reduce certain error/externality costs but increase latency and operational cost. Economic analysis must weigh avoided damages (legal, reputational, remediation) against increased transaction costs and slower throughput.
• Market structure and service design
  • Creates demand for third‑party verification, context-enrichment, attestation, and escrow services (markets for evidence providers, auditors, or “context verifiers”) that can satisfy required constraints cheaply.
  • Platforms might monetize faster-execute tiers by offering verified-context services or SLA-backed Right-to-Act clearance.
• Liability, insurance, and contracting
  • Hard constraints change legal risk allocation: demonstrable Right-to-Act checks provide stronger defenses in liability/neglect claims and could lower insurance premiums; conversely, failing to implement them may increase liability.
  • Contract design between platforms, users, and AI-vendors could incorporate explicit Right-to-Act clauses and penalties.
• Adoption incentives and moral hazard
  • Firms may under-invest in costly contextual checks if externalities are socialized (regulatory enforcement, reputational risk). Conversely, when firms internalize the benefits (fewer appeals, less remediation), they have incentive to adopt Right-to-Act systems.
  • Strategic actors may attempt to game constraint signals (e.g., falsified attestations). Economic models need to consider incentives for misreporting and build verifiability and audit mechanisms.
• Measurement and welfare consequences
  • New metrics needed: cost per blocked false-execution, benefits of prevented externalities, latency costs, and downstream welfare changes (user trust, churn).
  • Potential welfare gains from fewer wrongful irreversible actions (financial transfers, account suspensions) but efficiency losses from delayed legitimate actions.
• Regulatory and compliance economics
  • Regulators may prefer non-compensatory rules for high-stakes domains (finance, health, critical infrastructure). This influences compliance costs and creates barriers to entry that affect competition.
  • Certification markets could bifurcate: statistical/certification for aggregate behavior vs. Right-to-Act attestors for per-decision legitimacy.
• Implications for automation and labor
  • Hard per-decision checks may preserve human-in-the-loop roles (escalation queues), affecting labor demand and the pace of automation in regulated sectors.
  • Could shift the value proposition from full automation toward mixed systems where AI proposes and verifiers (human or automated) clear Right-to-Act constraints.
• Research directions for AI economics
  • Empirical studies: A/B tests on platforms comparing compensatory score pipelines vs. Right-to-Act constraints to estimate externality reductions, latency costs, and behavioral responses.
  • Mechanism design: design incentive-compatible verification protocols and pricing for context attestation services.
  • Insurance underwriting models: quantify loss distributions with and without Right-to-Act boundaries to price AI operational risk.
  • Market equilibrium models: analyze platform competition when some adopt Right-to-Act constraints and others do not (impact on trust, user adoption, litigation risk).
  • Estimation challenges: measuring recursive-data effects (certification trained on AI-shaped distributions) and modeling strategic feedback loops.
• Practical policy/economic prescriptions (derived)
  • For high-stakes domains, regulators and platforms should evaluate adopting non-compensatory Right-to-Act layers, possibly mandating minimal constraint classes.
  • Policymakers should consider standards for auditable non-action outcomes (defer/escalate logs) to lower monitoring and enforcement costs.
  • Economists and policymakers can develop cost-benefit frameworks to decide when hard constraints are socially optimal versus when scoring suffices.

Limitations and open questions relevant to economists - How to choose and cost constraints across heterogeneous decision classes? - Optimal calibration of which decisions require hard constraints vs. probabilistic scoring to balance social welfare and operational efficiency. - Strategic behavior and verification markets: how to design credible, low‑cost attestation to support Right-to-Act checks. - Empirical magnitude: how often do compensatory systems permit harmful executions that Right-to-Act would prevent, and what are the tangible economic losses avoided?

Overall, the paper supplies a crisp formal primitive (Right-to-Act) that reframes a design choice with clear economic trade-offs. For AI economics, it opens pathways to quantify the value of per-decision legitimacy, design markets and contracts around verification services, and study how hard constraints reshape incentives, adoption, and welfare in automated decision ecosystems.