Facts Need Passports: Governing Ethics, Privacy, and Bias in Knowledge Graphs

Knowledge graphs (KGs) are increasingly used as sources of knowledge for large language models (LLMs), hybrid AI systems, autonomous agents, and machine reasoning. However, conventional knowledge representation does not, by itself, establish whether a KG fact is ethically appropriate for a particular use, legitimately accessible to a particular agent, or endorsed by a relevant epistemic constituency. This paper proposes an EPB-passport framework that associates knowledge claims with explicit qualifications concerning Ethics, Privacy, and Bias (EPB). At the level of an RDF triple, the passport records whether a claim is appropriate for a specified agent or context, whether the agent is legitimately authorized to access or use it, and whether it is endorsed by a relevant epistemic constituency. These dimensions are distinguished from truth, provenance, and technical accessibility: endorsement does not establish truth, and technical retrievability does not establish legitimate access. The framework follows a conservative principle: a qualification must not be assumed merely because a claim exists or can be derived. Missing qualification information is treated as undetermined rather than as an implicit positive or negative judgment. The same principle applies to derived knowledge: an inferred claim cannot automatically inherit the EPB qualifications of its source facts but requires dimension-specific propagation and reassessment. The passport concept is also extended beyond RDF triples to semantic claims produced and manipulated by LLM-based systems, distinguishing computational units such as tokens and semantic spans from epistemic units such as claims and reasoning episodes. The paper develops and analyzes this conceptual and architectural framework, examining its implications for KGs, LLMs, LLM–KG hybrid architectures, and autonomous agents. Rather than reporting an implementation or empirical evaluation, it establishes a basis for explicit, context-sensitive governance of knowledge claims and identifies requirements for their formalization, propagation, and practical assessment.