Hugging Face study reveals producer-consumer rift in AI model supply chain

A new study from researchers at Queen's University and Université de Montréal published on arXiv (2607.21738) investigates the misalignment between model producers on Hugging Face and downstream consumers on GitHub. The team surveyed 50 Hugging Face producers and 95 GitHub consumers—the first dual-perspective analysis of its kind—across four AI supply chain dimensions: model discovery, documentation practices, lineage tracing, and model governance adoption. Their findings reveal a core disconnect: although both groups rely on the same documentation artifacts, they strongly disagree on where critical metadata should live. Producers tend to favor embedding metadata in model cards or repository READMEs, while consumers often expect structured metadata fields or standardized schemas that are missing in practice.

Lineage tracing also reveals divergent motivations. While 27.9% of producers and 31.4% of consumers reported tracing model lineage beyond the immediate parent model, producers primarily do so for provenance and reproducibility, whereas consumers are driven by quality and reliability concerns—seeking to verify the pedigree of models before integrating them into their software. On governance, producers believe that streamlining model release mechanisms would have the greatest positive impact, while consumers prioritize improved documentation and dependency transparency. These misalignments highlight persistent structural problems in the pre-trained language model (PTLM) supply chain, including inconsistent release practices, incomplete metadata, and divergence between Hugging Face and GitHub repositories. The study suggests that better documentation conventions, enhanced lineage visibility, and more consumer-aligned governance could reduce friction in model reuse and improve trust in the AI ecosystem.