A new governance framework argues AI deployments should only make claims supported by explicit, contextual evidence and must record the institutional constraints that measurement cannot override; RISE AI ties system-level evaluation to who owns power, what is non-negotiable (e.g., dignity), and where claims must stop.

Main Finding

AI systems must be evaluated not only as technical artifacts but as interventions into pre‑existing institutional and relational failures. Responsible deployment requires (a) an explicit rupture test that links the institutional baseline to system evaluation; (b) evidence‑bounded claims about what a system actually accomplishes; and (c) measurement‑bounded governance that records normative constraints that favorable evidence cannot override. The paper operationalizes this with RISE — an evidence architecture for producing bounded claims about Responsibility, Inclusivity, Safety, and Empowerment — and argues that measurement and design cannot by themselves settle deeper political‑economic or dignity questions.

Key Points

• AI reveals existing institutional weaknesses (responsiveness, belonging, accountability) and acts as an intervention that can repair, compound, substitute for, or conceal those weaknesses. Evaluation must address both revelation and intervention.
• Rupture test: before deployment, identify the prior institutional/relational failure, the human/non‑AI baseline option, and whether the AI repairs/compounds/substitutes/conceals that failure.
• From principles to protocols: the paper gives an operational chain — Principle → Design objective → System requirement → Implementation mechanism → Evaluation protocol → Bounded evidence claim.
• Design patterns (connect governance objectives to engineering mechanisms):
  • Answerability‑by‑Design (role definitions, provenance, logs, redress)
  • Contestability‑by‑Design (usable appeals, protection from retaliation)
  • Agency‑Preserving Interfaces (useful friction; retain authority)
  • Context‑Aware Evaluation (population/workflow/version matching)
  • Evidence‑Bounded Deployment (claims limited to collected evidence)
• Two fundamental limits of measurement:
  • Political economy/system boundaries: artifact‑level evaluation cannot by itself adjudicate ownership, financing, control, labor, or power asymmetries that shape legitimacy.
  • Dignity and construct validity: dignity is not a measurable latent variable; indicators can show violations of dignity but cannot measure dignity as such or trade it off against other outcomes.
• Measurement‑bounded governance: alongside evidence‑based claims, deployments should record explicit constraints (uses, populations, relationships) that no benchmark result can license or override.
• RISE architecture: produces versioned claim–evidence graphs linking bounded claims to indicators, evidence provenance, accountable owners, context, limitations, and expiry triggers; plus a context‑and‑power record (ownership, financing, labor, agenda‑setting) to make institutional order visible.

Data & Methods

• Methodological approach: conceptual and normative analysis combining:
  • Policy and standards review (EU AI Act, NIST AI RMF, ISO/IEC 42001)
  • Measurement and validation theory (construct validity emphasis: validity applies to inference)
  • Design pattern synthesis from governance and engineering literatures
  • Assurance and documentation practices (assurance cases, provenance, logs)
  • Ethical and political framing using Pope Leo XIV’s Magnifica Humanitas as a moral/political lens on dignity, distribution of technological goods, and political economy
• No primary empirical dataset: the paper is prescriptive/conceptual and constructs an operational architecture (RISE) for recording and assessing evidence and constraints.
• Operational artifacts proposed: versioned claim–evidence graphs, context‑and‑power records, and protocol templates for accountability, contestation, and monitoring.

Implications for AI Economics

• Make political economy explicit in economic assessments:
  • Evaluations must go beyond artifact performance to include ownership, control of data/compute, financing, labor inputs, and distribution of gains/losses. These institutional factors affect welfare, market structure, and bargaining power but are not captured by accuracy metrics.
• Incorporate rupture tests into cost‑benefit and impact analyses:
  • When assessing deployment benefits, compare outcomes to the relevant non‑AI baseline (what alternative institutions or human services would have provided) and account for substitution/externalization effects.
• Recognize limits of quantitative indicators:
  • Economists should not treat indicators (e.g., accuracy, uptake) as proxies for dignity, legitimacy, or social acceptability. Use indicators to support bounded causal inferences, but complement with normative, legal, and participatory inputs.
• Measurement‑bounded governance affects incentives and markets:
  • Explicit constraints (prohibited uses, protected populations) change allowable product scope and can shift investment incentives. Regulators and firms will internalize compliance costs; assurance markets (auditing, evidence‑provision services) will grow in importance.
• Valuation and procurement implications:
  • Public procurement and private investment should demand RISE‑style claim–evidence graphs and context‑and‑power records. This creates transaction costs but reduces asymmetric information about downstream social impacts.
• Distributional and labor effects:
  • Analysts must account for labor displacement, retraining costs, and who pays for institutional repair (public vs. private), including environmental/resource costs (compute, energy) that affect social welfare.
• Policy design and welfare modeling:
  • Welfare analyses should include non‑market values and institution‑level constraints (measurement‑bounded governance) as exogenous policy choices, not variables to be optimized purely by market outcomes.
• Empirical agenda:
  • Develop standardized indicators for bounded claims (e.g., measurable contestation outcomes, override rates, recourse effectiveness) and methods for linking these to economic outcomes (employment, access, consumer surplus).
  • Study how context‑and‑power features (ownership concentration, procurement terms) mediate realized benefits and distributional outcomes of AI deployments.
• Governance as an economic input:
  • Effective institutional capacity (e.g., trained reviewers, legal protections, monitoring) is itself an input with costs and returns; models should treat governance capacity as endogenous when predicting adoption and social outcomes.

Summary recommendation for economists: use the paper’s rupture test and RISE architecture to ground empirical and policy work in clearly specified baselines, bounded evidence claims, and explicit institutional constraints. This reduces the risk of conflating engineering performance with social legitimacy and helps align economic analysis with realistic, normatively informed assessments of AI’s social value.