Book

Description

Frontier artificial intelligence is governed through capability evaluations, risk thresholds, security controls and commitments to restrict or suspend systems whose risks cannot be adequately controlled. Yet a fundamental accountability gap remains: how can institutions prove that their predictions, safeguards and governance decisions remain effective when advanced AI systems produce real-world outcomes? Forecast-AI: Prediction-to-Outcome Accountability for Frontier AI Risk Governance presents an original multidisciplinary framework connecting what institutions predict before deployment with what they observe, decide and correct throughout the AI lifecycle. It establishes Prediction-to-Outcome Accountability as a continuing duty to maintain traceable relationships among capability forecasts, severe-risk predictions, safeguards, human authorisations, observed outcomes, discrepancies and corrective actions. Grounded in established safety, governance and accountability theories, FORECAST-AI integrates forecasting, monitoring, discrepancy detection, corrective action and verifiable evidence within a continuous governance cycle. Its non-compensatory approach prevents acceptable performance in one domain from concealing catastrophic failure in another. The framework distinguishes inaccurate forecasts, implementation failures, ineffective safeguards, deficient monitoring, changing contexts and delayed responses. It connects these failures to documented human review and proportionate decisions to continue, restrict, redesign or suspend an AI system. Written for regulators, policymakers, AI developers, independent evaluators, institutional leaders, governance professionals, investors and funders concerned with frontier-AI risk, this book provides a structured basis for assessing institutional claims, verifying safeguards, identifying governance failures, directing corrective action and informing responsible policy, investment and implementation decisions. As an original framework requiring context-specific validation before deployment, FORECAST-AI offers a rigorous pathway towards demonstrable institutional accountability for severe frontier-AI risks.