Open weights are not open source: Why AI's favorite label is under dispute
- ID
- 24577
- Status
- summarized
- Published
- 15 Sep 2026, 4:30 PM
- Fetched
- 15 Sep 2026, 5:16 PM
- Provider
- The Register
- Category
- technology
- Original URL
- https://www.theregister.com/columnists/2026/09/15/open-weights-are-not-open-source-why-ais-favorite-label-is-under-dispute/5295436
- Source URL
- https://www.theregister.com/headlines.atom
Summary
- Score
- 6.0
- Created
- 15 Sep 2026, 5:17 PM
- Tags
- Audience
- developersai_ml_learnerssaas_founders
What happened
Steven J. Vaughan-Nichols argues that AI companies routinely label models 'open source' when they only release open weights—the final trained parameters—without training data, architecture documentation, or reproducible pipelines. The Open Source Initiative and Stanford HAI's James Landay distinguish 'open weights' (you can deploy and fine-tune) from genuine 'open source' (you can inspect, reproduce, and understand why the model behaves as it does), calling the former 'open distribution' rather than openness.
Why it matters
If you self-host or fine-tune models from Hugging Face for production or client work, check what was actually published before assuming you have audit or redistribution rights. Weights alone let you run and fine-tune locally, but without training data or documentation you cannot verify provenance, reproduce results, or guarantee compliance—matters when regulators or enterprise customers ask what's inside your stack.
Discussion angle
When you pick a model for a Malaysian client or SaaS product, what's the practical risk of treating open-weights as open-source—does it matter for PDPA compliance, procurement requirements, or investor due diligence?