Scientific Knowledge Graph Construction Needs an AI-Mediated, Scientist-in-the-Loop Workflow (A Blue Sky Paper)

Scientific knowledge graph (KG) construction increasingly relies on large language models (LLMs) and heuristic pipelines to extract semantic structure from unstructured corpora. While automation enables scale, it often introduces ambiguity, overgeneralization, and inconsistencies—especially in biomedical and grey-literature domains. To address this, we propose a reflexive, AI-mediated, scientist-in-the-loop workflow that elevates human participation from labeling to semantic guidance. Unlike traditional HITL systems where humans serve as annotators or post hoc validators, our architecture enables the AI system to monitor confidence, detect failure modes, and initiate structured consultations with domain experts. Each intervention is triggered by statistical anomalies, graph discrepancies, or semantic drift, and is coupled with diagnostics, summaries, and provenance trails. We introduce the design of a modular enrichment pipeline augmented by diagnostic triggers, GGD-based corrections, and cube-based lineage tracking. This position paper articulates the motivation, architecture, and open challenges for realizing truly collaborative, transparent, and high-quality knowledge graph construction in science.