How Regulators Are Bringing AI Into Review (Without Losing Trust) with Maria Vassileva

What It Covers

Maria Vassileva, Chief Science and Regulatory Officer at DIA, examines how FDA and global regulators are integrating AI into drug and device review processes. Coverage spans FDA's ELSA tool deployment, validation frameworks, international agency progress, equity concerns, and the multi-stakeholder governance structures needed to sustain public trust.

Key Questions Answered

• •ELSA Validation Limits: FDA's internal AI assistant ELSA, deployed in early summer 2025, cannot be used in formal regulatory assessments due to documented hallucinations and false outputs. The lesson: validate the entire workflow process surrounding a model, not just the model itself, to prevent misplaced trust in partially validated tools.
• •Risk-Tiered AI Governance: FDA's January 2025 draft guidance establishes a risk-based credibility assessment framework where required scrutiny scales with context of use. Pair this with the Total Product Life Cycle approach and a Predetermined Change Control Plan, allowing manufacturers to pre-specify algorithm modifications and data updates without triggering full re-review.
• •MHRA Efficiency Benchmark: The UK's MHRA cut clinical trial approval times from 91 days to 41 days by integrating AI into regulatory review workflows. Their Innovation Office provides free, confidential regulatory advice to developers. This 55% reduction over roughly six months serves as a concrete performance target for other agencies pursuing similar reforms.
• •Bias Mitigation in Training Data: AI models trained on historical datasets can amplify existing racial, gender, and socioeconomic disparities in diagnosis and treatment decisions. Regulators now require developers to document data provenance, demonstrate model generalizability across patient populations, and build contextually scoped models rather than assuming bias can be fully eliminated from any single system.
• •Global Coordination Accelerating: PMDA Japan released a three-year AI action plan in October 2025, starting with document retrieval and translation before advancing to higher-risk applications in 2026–2027. DIA's AI Consortium currently includes six to seven regulatory agencies alongside industry and academia, building shared validation templates and taxonomies to reduce duplicated effort across jurisdictions.