AI governance and compliance

AI governance encompasses the policies, standards, and frameworks organizations and regulators use to manage risks associated with AI systems, including misinformation, bias, fairness, privacy, transparency, and accountability. As AI-generated synthetic media (text, images, audio, video) becomes increasingly realistic and easy to generate, governance frameworks address both the detection of AI-generated disinformation and the responsible deployment of AI systems that could enable such content.

Key frameworks and initiatives

MIT AI Risk Repository: A comprehensive inventory of AI risks across societal impact areas, including disinformation and misinformation. Used to structure evaluation protocols for benchmarks and risk scorecards.

EU AI Act (2024): The world's first comprehensive AI law, establishing a risk-based framework categorizing AI systems as prohibited, high-risk, limited-risk, or minimal-risk. Disinformation-generating systems fall under high-risk and may face mandatory algorithmic auditing and transparency requirements.

EU Code of Practice on Disinformation (2022 revision): Voluntary commitments by tech platforms and companies to combat disinformation through fact-checking, content labeling, advertiser transparency, and cooperation with researchers.

U.S. Executive Order on Safe, Secure, and Trustworthy AI (2023): Federal initiative setting standards for AI safety, AI-generated content labeling, and inter-agency coordination on AI risks.

Frontier Model Forum and Partnership on AI: Industry-led initiatives addressing frontier AI safety research, information sharing across government, academia, and industry.

Tech Accord on Combat Deceptive Use of AI in Elections (2024): Joint pledge by 20+ technology companies (Adobe, Amazon, Google, Meta, Microsoft, OpenAI, TikTok, X, etc.) to detect, label, and counter harmful AI-generated election content; commitments include rapid response and user education.

Applications to misinformation detection

Governance-aligned benchmarks and evaluation protocols serve several purposes:

1. Accountability: By explicitly mapping detection tasks to governance risk categories, researchers can assess whether systems mitigate identified harms.
2. Transparency: Rigorous evaluation metrics and public benchmark results enable stakeholders (regulators, platforms, civil society) to monitor AI-system performance.
3. Robustness testing: Governance frameworks often require evaluation under adversarial conditions (e.g., cross-modal attacks, distribution shift) to ensure systems fail gracefully under real-world deployment.
4. Risk scoring: Aggregating multiple metrics (accuracy, robustness, domain generalization) into risk indicators (e.g., disinformation risk, robustness risk) provides a holistic view of system reliability.

Key papers

• Raza et al. (2025) — VLDBench: Evaluating Multimodal Disinformation with Regulatory Alignment — Multimodal disinformation benchmark explicitly aligned with MIT AI Risk Repository; introduces Risk Scorecard mapping dataset components and evaluation metrics to governance risk areas (disinformation risk, robustness risk, domain shift risk); demonstrates integration of technical benchmarking with policy-relevant risk dimensions.

Related topics

• AI Safety — broader umbrella covering AI harms beyond misinformation
• Synthetic media — deepfakes, AI-generated text, and other AI-synthesized content
• Multimodal fake news detection — technical methods for detecting AI-generated multimodal content
• Disinformation — malicious false/misleading information
• Fact-checking and corrections — human-in-the-loop verification approaches

Notes

AI governance in the context of misinformation is an emerging area. Early frameworks (EU AI Act, US EO) are high-level and principles-based; practical benchmarks and operational standards are still developing. The VLDBench work is notable for explicitly operationalizing governance risk dimensions (from MIT Risk Repository) into measurable evaluation indicators, creating a bridge between policy requirements and technical benchmarking. Future work should extend this mapping to video, audio, and cross-lingual content, and develop audit and monitoring protocols for deployed systems.