JSON-LD 1.2 and Beyond

Extensions for Machine Learning Data Exchange

Abstract

JSON-LD has become the dominant format for structured data on the web, underpinning schema.org markup, Verifiable Credentials, and knowledge graph serialization. However, the rapid integration of machine learning into data pipelines exposes critical limitations: JSON-LD lacks native mechanisms to express prediction confidence, model provenance, temporal validity, or vector embeddings—metadata essential for trustworthy AI-to-AI and AI-to-human data exchange. Additionally, context injection attacks and unbounded recursion vulnerabilities pose security risks in production deployments.

This paper presents a systematic gap analysis of JSON-LD 1.1 against the requirements of modern AI systems, identifying 12 limitation categories spanning security vulnerabilities, performance bottlenecks, validation deficiencies, and data modeling constraints. We propose backward-compatible extensions addressing critical gaps: @integrity for hashlink-based context verification, context allowlist modes for restricting remote context loading, standardized resource limits, @confidence for quantifying prediction uncertainty, @source and @extractedAt for provenance tracking, @validFrom and @validUntil for temporal scoping, and a @vector container type enabling embeddings to coexist with symbolic knowledge graph data.

We validate the proposed extensions through implementation in a healthcare wearables context, demonstrating semantic interoperability between edge-based posture classification models and clinical knowledge systems. Our proposals align with the W3C JSON-LD Working Group’s current charter, establishing a foundation for representing AI-generated knowledge with appropriate epistemic humility and robust security guarantees.