Where are we at the moment with AI governance

The Partnership on AI (PAI) is a multistakeholder organization founded by industry, academia, and civil society to study and shape best practices for AI. Its publications synthesize technical, ethical, and policy insights and aim to produce actionable guidance that can be adopted by developers, deployers, and regulators. What PAI publishes and why it matters - Practical guidance: PAI issues reports, white papers, and toolkits on topics such as model cards, transparency, safety testing, red‑teaming, and risk assessment. These materials translate research and field experience into concrete practices organizations can implement. - Multistakeholder legitimacy: Because PAI brings together firms, researchers, and NGOs, its outputs carry cross‑sector credibility and help bridge differences between private incentives and public-interest goals. - Norm formation: PAI’s work shapes industry norms (soft law) by demonstrating feasible protocols for responsible development and deployment before—or alongside—formal regulation. - Community and capacity building: PAI runs working groups and convenings that surface use cases, share lessons from incidents, and incubate standards or prototypes (e.g., reporting templates, evaluation frameworks). - Policy input: PAI publications inform regulators and standards bodies by clarifying technical options and implementation tradeoffs (useful for policymakers drafting laws like the EU AI Act or for agencies designing oversight mechanisms). Model governance guidance — main emphases - Risk‑based approach: Prioritize resources and controls according to model capability and deployment risk (high‑risk systems require stronger safeguards). - Transparency and documentation: Encourage model cards, datasheets, and disclosure about training data provenance, evaluation metrics, and known limitations to support informed use and oversight. - Safety testing and red‑teaming: Advocate systematic adversarial testing, scenario analysis, and external evaluation to identify failures before deployment. - Human oversight and accountability: Recommend clear roles/responsibilities, auditability, incident reporting, and mechanisms for remediation when harms occur. - Privacy and security protections: Promote data governance, differential privacy, access controls, and measures to prevent misuse or leakage of sensitive information. - Continuous monitoring: Stress post‑deployment monitoring, feedback loops, and update/patch processes to address emergent issues. - Collaboration and information sharing: Support responsible sharing of vulnerabilities, best practices, and interoperable evaluation tools across actors. Representative PAI outputs - Model Card and Documentation guidance (templates and best practices) - Red‑teaming and adversarial testing reports - Governance frameworks and checklists for deployment and procurement (See Partnership on AI website for specific publications.) Why this guidance matters for governance PAI’s model governance guidance fills the gap between high‑level principles and operational practice. It helps organizations implement responsibilities that regulators may later require, provides policymakers with technically grounded options, and advances interoperable practices that can be incorporated into standards, procurement rules, and regulatory regimes. Selected sources: Partnership on AI publications and working groups; examples cited in OECD and EU policy discussions.