Model Card

Simple Definition

A model card is a short, structured document that an AI developer publishes alongside a model. It tells you what the model does, how it was trained, what it’s good at, what it struggles with, and what uses it wasn’t designed for.

The concept was introduced by Google researchers in 2018 as a way to make AI more transparent and accountable.

What a Model Card Contains

Most model cards cover:

Model description: what type of model it is, what it’s intended for
Training data: what data was used to train it, and from what time period
Intended uses: what tasks and contexts it’s designed for
Out-of-scope uses: what it shouldn’t be used for
Performance: benchmark scores across different tasks and demographic groups
Limitations: known weaknesses, failure modes, or biases
Ethical considerations: fairness concerns, risks, and mitigations
How to contact the developers: for reporting issues

Why Model Cards Matter

Before model cards, using an AI model meant trusting that it worked well without much information about how or why. Model cards give:

Users: a realistic picture of what the model can and can’t do
Developers: guidance on appropriate integration
Regulators and auditors: documentation for accountability
Researchers: a standard format for comparing models

Where to Find Model Cards

Hugging Face: most publicly shared models include a model card in their repository
Google: publishes model cards for models like Gemma and Gemini
Meta: publishes model cards for the Llama family
OpenAI: publishes system cards for GPT-5 and other models (sometimes called “system cards” or “safety cards”)
Anthropic: publishes model cards for Claude models

Limitations of Model Cards

Model cards are only as good as the information the developer chooses to share. They can be vague, incomplete, or promotional. They’re a step toward transparency, not a guarantee of it.