Evaluation as a tool for regulatory compliance: Scratching the AI regulation surface Carlos Muñoz Ferrandis, co-founder & COO

1278% 1000+ Policy initiatives reported by gvmts in 70+ jurisdictions in OECD database. May 2023. USA 2024: total number of AI-related regulations grew by Increase of AI policy and regulatory initiatives, worldwide AI mentioned in legislative proceedings 2022: 1,247 2023: 2,175 Between 2022 and 2023, AI incidents reported increased by globally... 56.3% Data extracted from OECD AI and AI Index Report 2024

Regulations are coming.

Main challenges... how to interpret regulation? how measure compliance? how to anticipate and mitigate risk? how to report compliance? ...Challenge for whom? Market Authorities

Prohibited Practices High Risk AI Systems General Purpose AI Evaluation & Red teaming Art 9, 15, 17 Art 53.1 Guardrails & Monitoring Art 9, 15, 17 Art 55.1 Documentation Art 11, 13, Annex IV Art 53, 55 EU AI Act control & risk mitigation tooling *Similarities with Digital Operational Resilience Act (arts 9, 10, 25)

EU AI Act Prohibited AI systems: “behavioral distortion” High risk AI systems: “robustness” “perform consistently for their intended purpose” Digital Operational Resilience Act Robustness, Resilience, Reliability of ICT systems Let´s scratch a bit the regulatory surface... How do we measure and monitor unclear requirements?

Define criterion “Robustness” Define metrics/rubrics Train “Evaluator models” Run evals, monitor at scale Close your eyes and pray Evaluation at scale is a need...and a science.

What is currently missing? Standardizing benchmarks for regulatory compliance Variety of open Benchmarks + not so good datasets Focus on defining criteria + metrics to measure Transversal criteria vs industry-specific criteria “Cards” everywhere Industry-specific Gen AI evals + guardrails

Safe & controlled deployment of gen AI Thank you!