What is high-stakes AI?

Definition

High-stakes AI describes AI systems whose outputs or actions can have serious consequences for people or organizations. Healthcare, employment, lending, legal decisions, critical infrastructure and public safety are common examples.

These systems require controls proportionate to their risks, including representative evaluation, expert oversight, traceability, secure data handling and clear human authority. Performance should be monitored in real use because standards, populations and model behavior can change after deployment.

ELI5

High-stakes AI is used where a mistake can seriously affect health, rights, safety, money or another important outcome. Examples include medical decisions, employment, lending, legal work and critical infrastructure.

For example, an AI that helps decide who receives treatment needs representative tests, expert oversight and a record of the evidence behind each recommendation. Monitoring must continue after release because people, standards and model behavior can change.

Frequently asked questions

What makes an AI use case high stakes?

A use case is high stakes when incorrect, biased or unavailable output can significantly affect health, safety, rights or essential opportunities.

How should high-stakes AI be governed?

It should use risk-based testing, qualified human oversight, audit trails, access controls, incident processes and continuous monitoring.

Videos explaining high-stakes AI

  1. Sebastian Fox beside the words Clinical AI Needs Better Evals