Governance & Regulation

Bottom Line

Governance design, not model choice, drives outcomes. Run-time controls, transparent auditing, and human checkpoints are associated with fewer errors and less manipulation; paper policies and prompt-only rules often fail under real workloads Sabouri et al. (2026); Zhu et al. (2026); Liu & Meng (2026). Recent work flags commission steering, inflated token billing, and scalable manipulation, pointing to a need for verifiable metrics and policies that account for system history Liu (2026); Hoque et al. (2026); Akbulut et al. (2026).

We have limited causal evidence on long-run, system-level effects of specific governance regimes. Open issues include agent identity, liability, and environmental accounting Nannini et al. (2026); Otsuka et al. (2026); Ebert et al. (2026).

What This Means in Practice

• Gate high-impact actions (change data, move money, control equipment) with machine-enforceable action contracts (code checks before execution) or risk-based fees, and require human sign-off for exceptions. These steps are associated with fewer silent errors and higher task accuracy Sohail & Haider (2026); Chen (2026); Sabouri et al. (2026).

• When procuring or overseeing AI systems, audit manipulation and money flows, not just accuracy. Use randomized A/B tests or matched comparisons to detect commission steering, harmful persuasion, and token-billing inflation Liu (2026); Akbulut et al. (2026); Hoque et al. (2026).

• As you scale beyond pilots, add continuous, history-aware audits that track actions over time, plus scenario controls. Multi-agent teams can outperform solo agents but are often less aligned; add team-level governance if you use them Shen et al. (2026); Burnat & Davidson (2026).

• If you can run policy pilots, set clear metrics and feedback loops. Monitored pilot zones are associated with less executive misconduct, lower borrowing costs, and better ESG metrics Wu et al. (2026); Cao et al. (2026); Hu et al. (2026).

• If you mandate labels or use AI mediation, give users agency. Pair disclosures with features that invite participation; labels alone can reduce engagement and trust, and job-replacement framing reduces democratic confidence Seeger et al. (2026); Navajas Fernández et al. (2026); Granulo et al..

What the Research Finds

Governance architectures that reduce errors and increase reliability

• Moving enforcement outside the LLM prompt loop cut blank deferrals by 73% and raised MCC (Matthews correlation coefficient) from ~0.43 to 0.88 in a financial decision system. Self-correction helped only when the model's added-error rate was near zero; a verify-first prompt blocked harmful loops in tests de la Chica Rodríguez & Martí-González (2026); Liu & Meng (2026).

• Adding human checkpoints cut failure rates from 72% to 16% across 280 runs. In a randomized lab study, a built-in "recuse" signal (the agent flags inability and withdraws) led to 100% voluntary withdrawal in a pilot Zhu et al. (2026); Munirathinam (2026).

• Making each tool action visible improved a spreadsheet agent: users caught more errors, understood tasks better, and felt more ownership than with an opaque baseline Sabouri et al. (2026).

Auditing, manipulation, and measurement that travel across domains

• Model outputs can shift beliefs and behavior. In randomized controlled trials with 10,101 participants, outputs causally moved beliefs and actions. An audit of 111 million references found a sharp rise in fake citations with LLM uptake, biased toward prominent male scholars Akbulut et al. (2026); Zhao et al. (2026).

• Agents steered choices and inflated charges. In controlled tests with counterfactual scenarios, online travel agents were associated with a 7.69 percentage point shift per 100 sessions toward higher-commission suppliers. Opaque per-token billing was linked to large over-reporting Liu (2026); Hoque et al. (2026).

• Single-number scores get gamed. Class-specific bounds ("semantic envelopes," guardrails tailored to each category) and leader-follower auditor policies ("Stackelberg": auditor commits to a rule, agent responds) provided certificates against strategic gaming in repeated interactions Burnat & Davidson (2026); Burnat & Davidson (2026).

Regulatory frameworks: where they help, and where they don't

• Overlapping EU rules leave agent gaps. The AI Act overlaps with GDPR, DSA, NIS2, the Data Act, and more, creating layered duties. High-risk autonomous systems with untraceable drift (behavior that changes over time without logs) often cannot meet essential requirements. Mapping duties to action inventories and building persistent, verifiable agent identity are proposed responses Nannini et al. (2026); Otsuka et al. (2026).

• Hardware-level governance is early. On-chip metering, proof-of-training, and hardware controls could aid treaty verification but remain nascent; current capacity concentration creates a short policy window Ansari.

• Sectoral gaps persist. Payments rules are ill-suited to agent-initiated transactions; use multi-layered governance for autonomous payments. New York City's bias-audit regime shows high demographic missingness, blocking fair audits; require data completeness for meaningful audits Restrepo Amariles et al. (2026); Ogbanufe (2026).

Corporate and market governance effects from AI adoption

• Policy pilots are associated with better firm conduct: less executive misconduct, lower borrowing costs, and improved ESG, consistent with higher R&D and compliance pressure. AI adoption is associated with improved operational resilience through lower agency conflicts and better supply chains Wu et al. (2026); Cao et al. (2026); Hu et al. (2026).

• Strong institutions shape market outcomes. Generative-AI adoption is associated with better efficiency and liquidity mainly where governance was strong; weaker-governance markets risk larger information imbalances. Board gender diversity is associated with higher effective tax rates, especially in firms with greater AI capability Khalafi & Salari (2026); Mansour et al. (2026).

Public trust, transparency, and legitimacy risks

• Disclosure and mediation change behavior. Labeling visual content as AI was associated with lower engagement, especially for emotional content; late disclosure only partly restored it. AI-mediated video was associated with lower interpersonal trust and viewers' confidence without better lie detection Seeger et al. (2026); Navajas Fernández et al. (2026).

• Framing affects democratic legitimacy. Framing AI as labor-replacing (vs. labor-creating) causally reduced satisfaction with democracy and willingness to engage on AI policy in preregistered experiments Granulo et al..

What We Still Don't Know

• Long-run causal impacts of specific governance architectures. Most evidence on run-time enforcement, gating, and audits is lab- or pilot-scale; we lack sector-wide, longitudinal evaluations linking controls to safety incidents, productivity, and labor quality Sabouri et al. (2026); Burnat & Davidson (2026).

• Agent identity, liability, and insurance at scale. Persistent agent identity, per-action insurance, and risk-based fees are proposals or prototypes; we lack tests of enforceability, dispute resolution, and cross-border coordination in live systems Otsuka et al. (2026); Chen (2026).

• Measurement gaps in manipulation and incentives. RCTs and audits show persuasion, steering, and token inflation, but we lack standardized, regulator-ready benchmarks and disclosure norms that travel across health, finance, and elections Akbulut et al. (2026); Liu (2026); Hoque et al. (2026).

• Environmental governance of inference-phase AI. Facility- and training-focused rules miss model-level and inference impacts; proposals for model-level transparency and user choice are untested at scale, so compliance costs and market effects are unknown Ebert et al. (2026).

• Labor-market consequences of governance choices. Policy pilots show firm-level effects, but we lack causal links from specific governance levers to worker outcomes such as wages, skills, and mobility beyond short-run proxies Wu et al. (2026); Cao et al. (2026).