Governance & Regulation

Bottom Line

Governance design often matters more than model choice in simulations; in firms and public systems, stronger governance is linked to lower misconduct risk and better resilience. Caveat: external validity—many findings come from simulations, reviews, or single jurisdictions, so generalizability to other settings and live systems is uncertain.

What This Means in Practice

• Make governance a launch gate: define enforceable rules, auditable logs, and human oversight for high-impact actions, and run pre-deployment stress tests with adversarial scenarios before granting real authority.
• Tie AI rollouts to corporate governance: strengthen internal controls and external monitoring, and review executive pay policies before major AI investments.
• Manage public expectations: avoid “AI will replace jobs” framing unless paired with policy support; design disclosure and user experience to preserve trust in high-touch services.
• Pair AI initiatives with enabling institutions: upgrade environmental disclosure rules, governance quality, and digital infrastructure to capture sustainability gains.
• Require audits that trigger fixes: mandate runtime guardrails and admission controls, and adopt standardized harmful-manipulation testing in development and before deployment.

What the Research Finds

Institutional design and runtime governance for AI agents

• In multi-agent simulations, changing governance structures (enforceable rules, auditable logs, human oversight) reduced corruption-like behaviors more than changing model identity; lightweight safeguards helped in some cases but did not reliably prevent severe failures Vedanta S P (2026).
• The same simulations recommend verifying integrity requirements pre-deployment and stress-testing systems under governance-like constraints before granting real-world authority Vedanta S P (2026).
• Prompts shape initial behavior but do not evaluate partial execution paths, motivating explicit runtime policies and checks Maurits Kaptein (2026).
• An admission-control specification (pre-approval checks for agent actions) enumerates 62 verifiable requirements and 12 prohibited behaviors as a baseline operational control set Marcelo Fernandez (2026).

Corporate governance and firm-level risk and performance

• In a firm-year panel, AI adoption was associated with less executive misconduct, alongside increased analyst and media scrutiny Jie Wu (2026).
• A Chinese AI pilot-zone policy increased operational resilience in treated firms, partly by reducing management agency conflicts Yiting Hu (2026).
• Stronger internal corporate governance weakened the increase in executive pay associated with AI adoption, consistent with limiting rent capture during technological change Jianan Shen (2026).

Public attitudes, legitimacy, and disclosure

• Preregistered experiments in the UK and US found that framing AI as labor-replacing rather than labor-creating reduced trust in democracy and willingness to engage politically with AI governance Armin Granulo.
• In a 37,079-person European survey, perceiving AI as labor-replacing was associated with lower democratic satisfaction and lower political engagement with technology after adjusting for covariates Armin Granulo.
• In an experiment on empathic communication, labeling identical responses as AI reduced perceived empathy relative to human-labeled or unlabeled replies Aakriti Kumar (2026).
• In a field experiment with an AI assigning work, personal experience with an AI boss did not measurably change attitudes toward AI in public decision making, but information exposure treatments produced significant attitudinal change even against prior dispositions Yotam Margalit.

Environmental and sectoral governance

• Most jurisdictions outside the EU lack AI-specific energy disclosure; current rules focus on facilities and training-phase emissions, not inference Kai Ebert.
• In a 104-country panel using econometric methods (GMM and 2SLS), stronger governance quality and better digital infrastructure mitigated the relationship between AI and CO2 emissions Partha Pratim Acharjee (2026).
• Across 450 observations, AI adoption correlated with higher energy justice scores, with effects strongest where environmental regulation is stricter and particularly on procedural justice Yong Ye (2026).
• In Indonesia’s health sector, policy and procurement are fragmented with weak transparency and explainability requirements for AI Wayan Sadwika.

Governance frameworks, audits, and evaluation tools

• A PRISMA-guided review (a structured systematic review) of 95 studies finds AI governance frameworks are fragmented, split between ethics/privacy and compliance/risk orientations, with persistent privacy–security tensions Orjuwan Albulayhi (2026).
• Practitioner research documents a results–actionability gap in LLM evaluations: teams get metrics but lack pathways or incentives to implement fixes Willem van der Maden (2026).
• A framework for evaluating harmful AI manipulation via human–AI interaction studies provides public testing protocols and materials to support uptake Canfer Akbulut (2026).
• In healthcare, technical advances such as synthetic data require adapted health technology assessment and governance pipelines, not standalone technical validation Ally Nyamawe.

What We Still Don't Know

• Whether governance designs that work in simulations generalize to live institutions remains untested at scale; current integrity evidence for agent collectives comes from controlled simulations, not production deployments Vedanta S P (2026).
• Beyond China’s pilot zones and global panel models, there are few natural or randomized experiments evaluating AI governance at subnational or sectoral levels, limiting causal inference about real-world effectiveness Yiting Hu (2026); Partha Pratim Acharjee (2026).
• Model-level energy and water footprints during inference are largely unreported outside the EU, so operational impacts remain poorly measured for policy targeting Kai Ebert.
• We lack randomized or natural-experiment evidence that linking audits to remediation reduces harms inside organizations; existing studies surface evaluation–remediation gaps without testing fixes Willem van der Maden (2026).
• How disclosure labels influence behavior and trust in high-stakes domains beyond empathy contexts is underexplored relative to their policy salience Aakriti Kumar (2026).