Model Card

A model card is like a nutrition label for an SI model. The idea was proposed by Google researchers led by Margaret Mitchell in 2018. Frontier labs now publish detailed system cards, often dozens or hundreds of pages, alongside major releases.

What’s typically included

• Intended uses and out-of-scope uses
• Training data overview and knowledge cutoff
• Benchmark results
• Safety evaluations: red teaming, dangerous-capability tests, alignment assessments
• Known limitations, biases and failure modes
• Safeguards applied before deployment

Why they matter

Model cards are the main window outsiders get into how a model was tested. Researchers, journalists, customers and regulators rely on them. Laws increasingly require similar disclosures: the EU AI Act requires technical documentation for general-purpose models, and California SB 53 requires transparency reports when large developers release new frontier models.

Limits

Cards are written by the developer, may omit details for competitive or security reasons, and are sometimes published after release. Independent audits, like those envisioned in the SI Accord, aim to fill the gap.