AI Risk Mitigation

Legal Framework

Article 9 of the AI Act establishes the core obligation for providers of high-risk AI systems to implement a risk management system as a fundamental component of their compliance architecture. The risk management system must be designed as a continuous, iterative process that runs throughout the entire lifecycle of an AI system — from development through deployment and post-market monitoring.

The framework requires providers to identify, analyze, and mitigate known and reasonably foreseeable risks that their AI systems may pose to health, safety, and fundamental rights. Critically, risk mitigation measures must be proportionate to the nature, scope, context, and purpose of the system, as well as to the severity and probability of potential harm. This proportionality principle mirrors the risk-based approach familiar from GDPR Article 9 processing conditions and Article 35 data protection impact assessments, where the adequacy of safeguards is measured against the objective risk landscape.

Recital 138 reinforces that risk mitigation is not a one-time exercise but must be embedded within regulatory sandboxes and controlled testing environments before systems reach the market. Recital 164 extends the supervisory dimension, granting the AI Office investigative powers to evaluate whether general-purpose AI model providers have implemented adequate mitigation measures, including through independent expert assessments.

The doctrinal analysis underscores that risk determination requires an objective assessment combining the probability of harm with the severity of its impact on rights and freedoms. This dual-factor analysis — probability multiplied by severity — sets the baseline for calibrating mitigation intensity.

Key Developments

The EDPB's SPE Programme report on AI Privacy Risks and Mitigations for Large Language Models, authored by Isabel Barberá, establishes a comprehensive risk management methodology specifically tailored to LLM systems. This represents one of the first concrete attempts to operationalize the AI Act's risk management requirements for a specific technology category, providing a structured approach to identifying and mitigating privacy-specific risks in generative AI.

In January 2026, the EDPB and EDPS jointly called for stronger safeguards to protect fundamental rights during AI Act implementation, signaling that supervisory authorities will scrutinize whether mitigation measures are substantive rather than procedural box-ticking. This positions risk mitigation as an area where enforcement bodies expect demonstrable, documented effectiveness — not merely the existence of a risk management policy.

Practical Guidance

• Implement a lifecycle-integrated risk management system that begins at the design phase and continues through post-market monitoring, documenting each iteration of risk identification, assessment, and mitigation in accordance with Article 9 AI Act requirements.

• Apply a dual-factor risk assessment methodology evaluating both the probability and severity of potential harm to fundamental rights, calibrating mitigation measures proportionately to the objective risk level rather than applying uniform controls across all systems.

• Document mitigation measure effectiveness through testing and evaluation, ensuring that residual risks are explicitly identified and justified — supervisory authorities expect evidence that measures actually reduce risk, not merely that measures exist.

• For LLM and generative AI deployments, adopt the structured methodology from the EDPB SPE Programme report, which addresses technology-specific risks including training data provenance, output filtering, and downstream misuse scenarios.

• Prepare for AI Office evaluations by maintaining accessible documentation of risk management decisions, mitigation measure selection rationale, and testing results, particularly for general-purpose AI models where independent expert assessment may be requested under Recital 164.