Architecting Candor: Products Liability and AI Incident Knowledge Governance

Artificial intelligence (AI) firms need incident knowledge to improve safety, yet the act of documenting that knowledge can increase litigation risk. As American courts increasingly subject AI systems to products liability, and as the European Union (EU)’s Product Liability Directive expressly classifies software as a “product,” the root-cause analysis necessary to diagnose an incident and remediate the system provides plaintiffs with evidence to establish fault and defective design. Such litigation risk produces a chilling effect and systematic underproduction of the formal incident knowledge that would otherwise drive safety engineering. It also deprives corporate boards of the incident data their oversight duties require. Moreover, as new law imposes incident-reporting duties, failures to preserve required records can create compliance risk, while the Product Liability Directive authorizes a rebuttable presumption of defectiveness when a defendant fails to comply with a court-ordered disclosure of relevant evidence. This playbook seeks to shift corporate AI incident response away from managing litigation risk and toward managing technical risk, showing that the two objectives, correctly structured, can coexist harmoniously. Drawing on the institutional designs through which aviation and healthcare resolved the same paradox, it proposes a three-channel architecture, a “Safety Translation Layer,” that separates automatically generated factual records from counsel-directed investigation and liability assessment while preserving a structured pathway through which objective safety signals become engineering requirements. A pre-committed telemetry tripwire, calibrated to the firm’s accumulated incident history, governs entry to the legal privilege channel. Grounded primarily in United States legal doctrine, the institutional architecture can be implemented today under existing law. In doing so, it aligns safer engineering and regulatory compliance with effective board oversight and a defensible litigation posture. The playbook concludes by recommending new legislation shielding organizations from liability or enforcement exposure related to reporting safety incidents.

Alumni

Meet the authors

(Research Team Lead)

Michael A. Celone

Michael A. Celone is a U.S. regulatory lawyer and a founder of the International AI Governance Lab at Harvard Law School. His work bridges law, policy, and technology to embed safety and fairness into AI systems ex ante.
Drawing on two decades of data-governance and machine-learning experience, he focuses on developing AI regulatory and governance frameworks that align innovation with public trust and transparency in ways that advance human dignity and societal good.

Mosi Secret

Mosi is an award-winning investigative journalist with over fifteen years' experience reporting for The New York Times Magazine, ProPublica, GQ, This American Life, and National Geographic, and host of the iHeart/Campside Media podcast Radical. He is currently under contract with Little, Brown on a nonfiction book about the desegregation of Southern boarding schools.

Mosi is now applying his investigative skills to AI governance research with Arcadia Impact, and is seeking roles in AI policy, research, and communications.

Noga Bregman

Noga is an AI researcher who builds and trains AI agents.

As an early member of the startup Tzafon AI, she worked across agent development, tool use, reinforcement learning, and continuous learning, and built evaluation environments to measure agent performance - helping grow the team from four people to twenty.

She has also worked in PayPal's horizontal AI department and previously researched algorithmic governance.

Bekhzodkhon (Beck) Alikhanov

Beck is a strategist with 12+ years across sovereign investment, M&A, and corporate transformation, including roles at SQB Bank, the UK Department for Business and Trade and UzAssets ($1.5B sovereign fund).

A Harvard Kennedy School MPA, he pairs executive judgment with hands-on AI fluency.

At Arcadia, he researches how safety-reporting regimes from aviation, healthcare, and pharma can help AI firms move incident knowledge to engineers without amplifying legal risk.

Eduardo Mignot

Eduardo Mignot is an AI product leader with 12+ years building and deploying AI across regulated European industries; open banking (PSD2), fintech, and insurtech; including LLM-powered features shipped to thousands of users.

On Arcadia Impact's AI Governance Taskforce, he researches why AI companies miss their own internal warning signs: how safety signals get lost between engineers, lawyers, and the board.

He is now seeking AI governance roles where deployment experience can make oversight frameworks workable in practice.

Alumni

Meet the authors

Michael A. Celone

(Research Team Lead)

Michael A. Celone is a U.S. regulatory lawyer and a founder of the International AI Governance Lab at Harvard Law School. His work bridges law, policy, and technology to embed safety and fairness into AI systems ex ante.
Drawing on two decades of data-governance and machine-learning experience, he focuses on developing AI regulatory and governance frameworks that align innovation with public trust and transparency in ways that advance human dignity and societal good.

Mosi Secret

Mosi is an award-winning investigative journalist with over fifteen years' experience reporting for The New York Times Magazine, ProPublica, GQ, This American Life, and National Geographic, and host of the iHeart/Campside Media podcast Radical. He is currently under contract with Little, Brown on a nonfiction book about the desegregation of Southern boarding schools.

Mosi is now applying his investigative skills to AI governance research with Arcadia Impact, and is seeking roles in AI policy, research, and communications.

Noga Bregman

Noga is an AI researcher who builds and trains AI agents.

As an early member of the startup Tzafon AI, she worked across agent development, tool use, reinforcement learning, and continuous learning, and built evaluation environments to measure agent performance - helping grow the team from four people to twenty.

She has also worked in PayPal's horizontal AI department and previously researched algorithmic governance.

Bekhzodkhon (Beck) Alikhanov

Beck is a strategist with 12+ years across sovereign investment, M&A, and corporate transformation, including roles at SQB Bank, the UK Department for Business and Trade and UzAssets ($1.5B sovereign fund).

A Harvard Kennedy School MPA, he pairs executive judgment with hands-on AI fluency.

At Arcadia, he researches how safety-reporting regimes from aviation, healthcare, and pharma can help AI firms move incident knowledge to engineers without amplifying legal risk.

Eduardo Mignot

Eduardo Mignot is an AI product leader with 12+ years building and deploying AI across regulated European industries; open banking (PSD2), fintech, and insurtech; including LLM-powered features shipped to thousands of users.

On Arcadia Impact's AI Governance Taskforce, he researches why AI companies miss their own internal warning signs: how safety signals get lost between engineers, lawyers, and the board.

He is now seeking AI governance roles where deployment experience can make oversight frameworks workable in practice.

Programme

AI Governance Taskforce

The AI Governance Taskforce is a career development programme for experienced professionals looking to transition careers into AI governance, focussed on reducing risks from advanced AI.
Participants work around existing commitments during our 12 week, remote, part-time cohorts, producing policy research in teams of 4, led by our Research Team Lead staff in partnership with recognised experts in the field. Teams write an academic-style paper and accompanying blog post to build knowledge, skills and work portfolios.