Value: A Verifiable Alignment Layer for Agentic AI (asteris.ai)

The recently proposed VALUE framework introduces a novel independent alignment layer designed to ensure AI agents' actions adhere to clearly defined value systems. This framework addresses the concern that as AI agents grow increasingly capable—managing complex tasks and making autonomous decisions—the risk rises that they might take inappropriate or harmful actions despite fulfilling their goals. The VALUE layer evaluates proposed actions against established value principles, producing verifiable evidence before such actions are executed. It distinguishes between goal alignment (accomplishing a set objective) and value alignment (adhering to ethical or moral standards), emphasizing the latter's importance as AI capabilities advance. VALUE employs a tiered review process through components like the VALUE Broker, which mediates the authority for actions, and VALUE Attestation Objects (VAO), which document the evaluations and their contexts. This framework acknowledges the diversity of value systems across organizations and builds mechanisms for consensus-based evaluations, potentially allowing multiple value authorities to weigh in on each action. The architecture aims not only to enhance safety and accountability in AI systems but also to establish a rigorous framework for AI governance amid the escalating capabilities of these technologies, thereby addressing critical issues of trust and oversight within the AI/ML community.