5 exercises — Learn AI governance and policy vocabulary: AI Safety Institute, voluntary commitments, frontier AI, capability thresholds, red-teaming requirements, and safety evaluations.
0 / 5 completed
1 / 5
What is an AI Safety Institute (AISI) in the context of AI governance?
AI Safety Institutes (e.g. the UK AISI, US AISI) are government-affiliated bodies whose mandate is to evaluate advanced AI models for safety risks — including dangerous capabilities, misuse potential, and systemic risks. They conduct pre-deployment testing and publish safety evaluations to inform regulation.
2 / 5
In AI governance, a voluntary commitment by an AI company typically means:
Voluntary commitments (e.g. the White House July 2023 commitments, the Seoul AI Safety Summit pledges) are self-regulatory pledges by AI companies — covering practices like pre-deployment red-teaming, information sharing with governments, and safety evaluations. They carry no legal penalties for non-compliance but create reputational and political accountability.
3 / 5
What is a frontier AI model in the context of AI policy and regulation?
"Frontier AI" refers to the most advanced AI models currently being developed — those at or near the capability frontier. Regulatory frameworks (e.g. the EU AI Act, UK AI Safety Summit) use this term to designate the AI systems that pose the greatest potential risk and therefore require heightened scrutiny, evaluation, and safety obligations.
4 / 5
A policy document states: "The model must pass safety evals before deployment." What does this mean in practice?
"Safety evals" refers to structured evaluations designed to test whether a model has dangerous capabilities — such as providing meaningful uplift to those seeking to create biological, chemical, nuclear, or radiological weapons, or assisting cyberattacks. Companies like Anthropic publish "responsible scaling policies" specifying capability thresholds that trigger deployment restrictions.
5 / 5
What does red-teaming mean specifically in an AI governance / pre-deployment context?
In AI governance, red-teaming means deliberately trying to make the AI system fail in harmful ways — eliciting dangerous information, bypassing safety guardrails, generating CSAM, or providing uplift for weapons. Policy frameworks increasingly require documented red-teaming before frontier model deployment, with results shared with relevant government bodies.