Model Card in plain English.

A model card is a standardized document a lab publishes alongside a new AI model, summarizing what it was trained on, what it's good and bad at, its benchmark scores, known failure modes, and safety testing results. The concept was proposed by researchers in 2018 as a "nutrition label" for AI models, and by 2026 publishing one is close to standard practice for any major model release — Anthropic calls its version a "system card," OpenAI publishes similar documents for GPT releases. Model cards are useful for developers deciding which model fits their use case and for researchers tracking how capability and safety testing evolve release over release. Their biggest limitation is that labs write and publish their own model cards, with no external audit requirement in most jurisdictions, so the level of detail and candor varies considerably between companies and even between releases from the same company.