Model Disclosure Requirements

Until ~2023, what a lab disclosed about a new model was entirely voluntary. You got a blog post, maybe a technical report. Weights, training data, evaluation methodology, and safety testing were proprietary. That norm is changing, unevenly.

Formal requirements (as of ~2025-2026)

• EU AI Act: General-purpose model providers must publish detailed documentation and copyright-policy summaries; systemic-risk models face additional obligations including evaluations and incident reporting
• UK AI Safety Institute pre-deployment evaluation access (voluntary but expected)
• California SB 53 (signed 2025): transparency reports from frontier developers on safety frameworks
• White House voluntary commitments (2023): red-teaming disclosures and content-provenance commitments from major labs

What's still mostly opaque

1. 1Exact training data composition — most labs share only high-level descriptions
2. 2Red-team findings — summarized at best, rarely reproducible
3. 3Copyright-relevant data sources — a subject of active litigation
4. 4Fine-tuning and alignment dataset specifics
5. 5Deployment monitoring data

“Companies that publish the most detailed safety information tend to be the ones doing the most safety work. That correlation could reverse if disclosure becomes mandatory and checkbox-shaped.”

The big idea: disclosure is the first step toward accountability. Accountability requires someone qualified to read the disclosure and act on it. Both are still under construction.