Skip to main content
TPDI logo
Back to Layer 3

Capability 3.1.5

Safety & Value Alignment

Definition

Models, or model ecosystems with multiple AI models, tools or agents, that comply with local ethics principles, privacy and safety regulations and societal expectations. This may entail developing an AI fabric with embedded guardrails as code.

Examples

  • Using techniques such as reinforcement learning to align pre-trained models with Australia's AI Ethics Principles.

Indicators

  • Number of domestic models incorporating ethical or safety tuning
  • Existence of transparency and audit frameworks
  • Public red-teaming or risk reports

National AI Plan

PLAN COMMITMENT

Moderate commitment

Prioritises safe AI development and deployment through establishment of the AI Safety Institute (AISI), assurance guidance and risk management initiatives (Action 7); also prioritises AI in line with democratic values (Action 9). Given the value of committed funding for AISI (as compared to other jurisdictions), commitment is assessed as moderate. Warrants attention to identify data (given this capability could generate very high agency and leverage). It is particularly important to increase agency in this capability if investments are not made in General Purpose & Frontier Models (3.1.1.10).

Stocktake

Is the Stocktake missing evidence?

Share additional information for potential future Assessments