| { |
| "field_id": "F016", |
| "field": "ai safety", |
| "directions": [ |
| { |
| "direction_id": "G0", |
| "name": "Societal-scale AI risk pathways and governance", |
| "scope": "Study how AI capabilities, actors, and failures can give rise to catastrophic or other major societal risks, and study institutional, release, and access arrangements that limit high-risk misuse and extreme risks. Risk-pathway analysis and governance arrangements are both within scope; a study need not address both." |
| }, |
| { |
| "direction_id": "G1", |
| "name": "Harmful generative output and safeguard resilience", |
| "scope": "Study when generative systems produce harmful content for users, and how to discover safeguard vulnerabilities or change or constrain generation behavior to reduce such outputs. Vulnerability probing and output prevention are each within scope; neither is a prerequisite for the other." |
| }, |
| { |
| "direction_id": "G2", |
| "name": "LLM application trust-boundary security", |
| "scope": "Study vulnerabilities and defenses in LLM-integrated applications when external content, application data, or tool interactions cross the trust boundary between instructions and data." |
| }, |
| { |
| "direction_id": "G3", |
| "name": "AI safety assessment and assurance", |
| "scope": "Study how to identify, measure, document, verify, or treat AI-system safety risks and provide testable safety grounds for use or deployment. Risk characterization of an individual model and lifecycle hazard management are both within scope." |
| }, |
| { |
| "direction_id": "G4", |
| "name": "Specification and fidelity of intended AI goals", |
| "scope": "Study which human goals, values, or behavioral rules AI should follow, and how learning or optimization processes acquire and preserve those intended goals. Specifying goals and preventing goal deviation are related branches; a study need not address both." |
| }, |
| { |
| "direction_id": "G5", |
| "name": "Oversight and human control of strategic AI", |
| "scope": "Study how more capable or potentially strategic AI can be subject to limited human oversight, and how to identify and constrain behavior that would undermine correction, shutdown, or other forms of human control." |
| }, |
| { |
| "direction_id": "G6", |
| "name": "Deceptive AI influence on users", |
| "scope": "Study how user-facing AI persuasion or deception distorts people's beliefs, judgments, or autonomy, and how to identify misleading statements." |
| } |
| ] |
| } |
|
|
