Advancing Responsible AI Innovation A Playbook 2025

Government leaders Key roadblocks organizations encounter from the broader ecosystem Limited incentives to report on responsible AI practices, discouraging companies from sharing information for fear of facing reputational and liability issues Lack of standardized incident reporting protocols, impeding the collection of reliable and comprehensive data, critical for preventing and mitigating future incidents Opacity of AI’s environmental impact: 84% of generative AI use is done through undisclosed models.61 As AI adoption grows, data on the environmental impacts is increasingly scarce, fuelling misinformation and public misconceptions. Ethics washing occurs when companies overstate their capabilities in responsible AI, creating an uneven playing field where genuine efforts are discouraged or overshadowed by exaggerated claims. Static benchmarks becoming misaligned with emerging risks, especially as many popular benchmarks are reaching saturation points or suffer from a lack of transparency, reproducibility and real-world relevance Actions for government leaders –Assess the state of responsible AI practices in the industry: Policy-makers must understand the state of responsible practices by AI providers and industry users within their jurisdiction. Such assessments can: –Incentivize organizations to measure maturity: Build awareness of the actual state of responsible AI practices within the organization. –Support evidence-based policy: Educate policy-makers on industry practices and responsible AI implementation challenges. –Prevent unnecessary regulation: In cases where enough companies demonstrate proactive and sufficient risk management. –Provide insights into forthcoming AI capabilities: Stay abreast of developing AI to assess potential opportunities and challenges of jurisdictional interest, such as national security. Jurisdictions should consider the advantages and limitations of various reporting instruments when incentivizing industry to report responsible AI practices (see Table 2). Layering multiple instruments can help offset trade-offs and bolster overall efficacy (see Case study 7). Mandating reporting in select instances may offset participation challenges. Additionally, governments should support academia, civil society, and third-party efforts to assess the state of responsible AI practices. –Standardize and incentivize risk and incident reporting: Promote compliance, data quality and insights gathering across jurisdictions through harmonized taxonomies, safe harbour provisions, and interoperable disclosure platforms that encourage transparency while safeguarding innovation. The level of disclosure for risks and incidents may vary depending on the audience and availability of expertise and resources required to analyse information.62 Disclosures must balance data privacy and security, particularly when reporting incidents related to vulnerable populations. Participation incentives for reporting could include access to other organizations’ reported incidents or mandated disclosure. For example, the EU AI Act requires general-purpose AI model providers of high-risk systems to “track, document and report relevant information about serious incidents and possible corrective measures to address them.”63 –Drive the evolution of benchmarks and standardize validation: Access to updated code, test sets and validation methods is needed to ensure companies and regulators base decisions on accurate metrics for system performance. Convene industry, academia and civil society to identify benchmarks and standards for AI safety assessments across industries and contexts. For example, Singapore’s Global AI Assurance Pilot gathered 17 organizations from 10 industries and nine countries to co-develop norms and practices for generative AI testing.64 –Facilitate environmental transparency disclosures: Considerations must be given to voluntary or mandatory industry-wide measures that include publishing impact data (e.g. energy, carbon, water) across the AI value chain, integrating AI’s environmental costs into corporate reporting and procurement, and developing standardized verification processes (see Case study 2).65 Advancing Responsible AI Innovation: A Playbook 25