Summary
- Nineteen UK computer scientists want an independent body to investigate serious AI incidents and near misses.
- The proposal draws on accident investigation models used in safety critical industries.
- The researchers also want human override mechanisms, clearer accountability, and stronger government procurement standards.
Nineteen UK computer scientists have called on the government to establish an independent authority capable of investigating serious artificial intelligence incidents and near misses, using an institutional model comparable to accident investigation in aviation.
Under the proposal, investigators would examine what happened when advanced AI systems behave dangerously, publish findings, and use evidence from individual cases to improve future safety practice without requiring every investigation to begin as an enforcement action.
Accident investigation bodies in transport typically reconstruct events and identify contributing factors separately from regulators or prosecutors responsible for sanctions, allowing technical and organisational failures to be examined even where responsibility is distributed among several parties.
Applying that distinction to AI could become more relevant as autonomous systems navigate websites, operate software, execute code, and complete chains of actions with limited direct supervision, since a harmful outcome may emerge from interactions between a model, tools, permissions, data, and human instructions.
Autonomy complicates accountability
Where an AI agent combines technology from several suppliers, the question of responsibility can extend across the model provider, organisation deploying the system, external tools it invokes, people who configure its permissions, and the person who initiated the task.
The researchers have consequently paired their call for independent investigations with proposals for human override mechanisms and clearer legal and organisational accountability when autonomous systems cause harm.
Government procurement forms another part of the proposal, with the signatories arguing that public bodies buying advanced AI systems should require identifiable parties to remain legally accountable rather than allowing responsibility to fragment across suppliers and technical components.
Recent security incidents involving agents have given those questions a practical rather than purely theoretical basis. Cyber Insider has reported on research agents reaching systems they were not intended to access and on autonomous tools interacting with external infrastructure in unexpected ways.
Existing UK institutions already conduct technical research into AI safety and security, while regulators retain authority within areas such as data protection, competition, and sector specific oversight. A dedicated investigation body would occupy a different position by examining individual failures after they occur and publishing lessons from them.
Whether government adopts that model will determine whether serious failures involving autonomous AI acquire a dedicated investigative route or continue to be divided among technology suppliers, existing regulators, affected organisations, and research bodies.
