Only 7 percent of open model licenses survive to production

Researchers traced 232,270 dataset-to-model-to-application chains across Hugging Face and GitHub to check whether license obligations actually make it downstream. They mostly don't. Every obligation-bearing license category ends up below 7% end-to-end survival, per the paper (arXiv 2607.20300). Permissive licenses, by contrast, survive at 95.1%.

The failure mode isn't malice, it's erosion. 62.3% of chains pass through at least one artifact with no declared license at all, and that gap concentrates in a small set of foundational datasets that everything downstream depends on. One undocumented link early in the chain and every obligation attached to it, attribution, share-alike, non-commercial, whatever, just disappears by the time someone ships an app.

This matters because most teams treat the license field on a Hugging Face model card as ground truth. It isn't. It's whatever the last person in the chain typed in, often without checking what the dataset or base model actually required. If you fine-tuned on top of someone else's fine-tune, you inherited a license that may have already been wrong three hops back.

I've pulled models off Hugging Face for side projects without ever tracing what dataset trained the base checkpoint. I'd guess most people building on open models do the same. The paper's number gives that habit a size: a 93%+ failure rate on obligation-bearing licenses isn't a rounding error, it's the default outcome.

What this doesn't tell you is which specific models in your stack are affected, the study measures the ecosystem, not your repo. If you're shipping something built on a fine-tuned model with any license other than permissive, the obligation is on you to trace the lineage yourself, not to trust the card.