AI Usage Cards vs Model Cards
Model Cards document AI models themselves while AI Usage Cards document how researchers use those models. Learn when you need which framework.
Two Sides of AI Transparency
If you have been looking into AI documentation frameworks, you have probably come across both Model Cards and AI Usage Cards. They sound similar, and both aim to increase transparency around AI. But they answer fundamentally different questions.
Model Cards answer the question, "What can this model do, and what are its limitations?"
AI Usage Cards answer the question, "How did you use AI in this specific project?"
These two frameworks are complementary. They were designed for different audiences, created by different people, and serve different purposes in the AI transparency ecosystem. Understanding the difference will help you figure out which one you actually need for your situation.
What Are Model Cards?
Model Cards were introduced by Mitchell et al. at Google in 2019 in their paper "Model Cards for Model Reporting." The idea was straightforward. When a team releases a machine learning model, they should provide a standardized document that describes the model's performance characteristics, intended use cases, limitations, and ethical considerations.
Think of a Model Card as a nutrition label for a machine learning model. It tells you what is inside, what it is good for, and where it might fall short. Google publishes Model Cards for many of its own models, and the practice has spread across the AI industry. Hugging Face, for example, encourages Model Cards for every model hosted on its platform.
A typical Model Card includes information about the model's architecture, training data, evaluation metrics across different demographic groups, intended uses, and out-of-scope uses. The people who fill out a Model Card are the developers and researchers who built the model.
What Are AI Usage Cards?
AI Usage Cards were introduced by Liebrenz et al. at JCDL 2023 (DOI: 10.1109/JCDL57899.2023.00060) as a way for researchers to document how they used AI tools in their scientific work. The motivation came from a growing problem in academia. More and more researchers were using tools like GPT-4, GitHub Copilot, or Midjourney in their research workflows, but there was no standardized way to report that usage.
An AI Usage Card captures which AI tools were used, what tasks they performed, how their outputs were verified, and what role the AI played in the research process. The person who fills out an AI Usage Card is the researcher who used the AI, not the company that built it.
You can generate an AI Usage Card for free at ai-cards.org.
Head-to-Head Comparison
| Dimension | Model Cards | AI Usage Cards |
|---|---|---|
| Purpose | Document an AI model's capabilities, limitations, and performance | Document how AI tools were used in a specific research project |
| Introduced by | Mitchell et al. at Google (2019) | Liebrenz et al. at University of Goettingen (JCDL 2023) |
| Primary audience | Model users, regulators, downstream developers | Journal reviewers, readers, fellow researchers |
| Who fills it out | The team that built or trained the model | The researcher who used the AI tool |
| Scope | One model, evaluated across different conditions | One research project or paper |
| Typical format | Structured document with sections on performance, fairness, and intended use | Structured card with fields for tool, task, verification, and human oversight |
| Key fields | Model details, intended use, performance metrics, limitations, ethical considerations | AI tool used, purpose of use, output verification method, degree of human oversight |
| When you need it | When releasing or deploying a model | When submitting a paper or reporting research that involved AI |
When Do You Need Which?
You need a Model Card if you are releasing a machine learning model for others to use. If your team has trained a new language model, image classifier, or recommendation system and you plan to make it available, a Model Card helps downstream users understand what they are getting. Companies like Google, Meta, and OpenAI publish Model Cards (or similar documents) alongside their model releases.
You need an AI Usage Card if you are a researcher who used AI tools during your work and want to transparently report that usage. If you used ChatGPT to help draft sections of your paper, or used Copilot to generate analysis code, or used DALL-E to create figures, an AI Usage Card is the right framework. It tells your readers exactly what role AI played and how you validated its contributions.
You might need both in some cases. Imagine you are a researcher who trained a new model and also used GPT-4 to help write the training pipeline code. The new model you are releasing would benefit from a Model Card. Your paper describing the work would benefit from an AI Usage Card that documents how GPT-4 helped with the code.
A Practical Example
Consider Dr. Amara, a computational linguist. She fine-tuned a BERT variant for detecting sarcasm in social media posts. During the project, she used GitHub Copilot to help write her data preprocessing scripts and used ChatGPT to brainstorm evaluation metrics.
For the model she is releasing, she creates a Model Card that documents the model's accuracy across different social media platforms, its performance on different demographic groups, known failure cases (for example, it struggles with sarcasm in languages other than English), and intended uses.
For the paper she is submitting to a journal, she creates an AI Usage Card that documents her use of GitHub Copilot for code generation and ChatGPT for brainstorming. She notes how she reviewed all Copilot suggestions manually and that the evaluation metrics suggested by ChatGPT were validated against the existing literature.
They Work Together, Not Against Each Other
The AI documentation ecosystem works best when different frameworks cover different aspects of transparency. Model Cards tell you about the tools. AI Usage Cards tell you how those tools were applied. Together, they give a much fuller picture.
For a broader look at how these frameworks relate to Datasheets for Datasets, System Cards, FactSheets, and other documentation approaches, see our full comparison of AI documentation frameworks.
If you are interested in how AI Usage Cards compare to other specific frameworks, take a look at AI Usage Cards vs System Cards or AI Usage Cards vs Datasheets for Datasets.
Generate Your AI Usage Report
Create a standardized AI Usage Card for your research paper in minutes. Free and open source.
Create Your AI Usage Card