What is a model card? Explained with the label on a food packet
A model card is the document published with an AI model that says what it is for, how it was tested and where it falls short. Think of it as the label on a food packet.
- Author
Umair Mateen- Published
- Reading time
- 5 min
- Category
- AI
When Anthropic released Claude Fable 5.2 this week, it published a full system card alongside it (our news post), and OpenAI published one for GPT-6 Cyber last week. Here is what those documents are, and why they are worth a look before you pick a model.
In one sentence: A model card is a document published with an AI model that explains what the model is for, how it was tested, how well it did, and where it falls short, so people can judge whether it suits their job before they use it.
The everyday picture
Pick up a packet of food in any shop and turn it over. The label tells you what is inside, who made it, what it is meant for and what to watch out for, such as allergens. It also lists measured facts, like the nutrition values per serving.
You do not need to visit the factory to decide whether to buy it. The label gives you enough to compare two packets on the shelf and to spot the one that is wrong for you.
A model card is that label for an AI model. The company that built the model writes it, and you read it before you put the model to work.
The maker tests the model and prints the results on its card. You read the card, then decide whether the model suits your job.
How it works, step by step
- The maker builds the model (the factory makes the food). The company decides what the model is meant to do and what data it learned from.
- The maker tests it (the food is measured). Teams run evaluations: sets of test tasks that measure skills such as coding or reasoning, and safety checks that try to make the model misbehave. Some makers also invite outside testers.
- The results go on the card (the label is printed). A typical model card covers intended uses, known limitations and biases, the training data, and evaluation results.
- You read the card (you turn the packet over). You check whether the model was tested on work like yours, and what it is not meant for.
- You decide (you buy, or put it back). The card helps you choose a model, set safeguards around it, or plan your own tests.
Try it · Myth or fact
Model cards: myth or fact?
Decide for yourself first, then tap each statement to check it.
0 of 3 checked
Where the picture stops
- Food labels follow a fixed layout that is the same on every packet. Model cards do not: each company chooses its own format and depth, from a short web page to a long report.
- Most of the numbers on a model card come from the maker's own tests. Treat them as the maker's claims, not as independent proof that the model will do well on your work.
- Some companies call their version a system card. Anthropic says its system cards document a model's capabilities, safety evaluations and deployment decisions, so they go beyond the model's skills to how it is released and safeguarded.
Why it matters for your business
- Fewer bad surprises. The limitations section tells you up front where a model is weak, before your customers find out.
- Faster comparisons. Cards let you compare models on the same kinds of tests without running everything yourself first.
- Lower risk. Cards describe safety testing and what a model should not be used for, which helps you decide what checks and human approvals to add.
- Easier sign-off. Security, legal and compliance teams often ask what a model is and how it was tested. The card is the first place to point them.
An example
Imagine a 30-person travel agency choosing an AI model to draft replies to customer emails in three languages. Before signing up, the manager reads the model cards of two options. One card shows test results in all three languages. The other only reports results in English. The manager shortlists the first, then runs 50 of the agency's own past emails through it to check the cards' claims on real work.
Frequently asked questions
Who writes a model card?
Usually the company or research team that built the model. On open model sites such as Hugging Face, every model page has a card, kept as a README file alongside the model.
What is the difference between a model card and a system card?
They cover the same ground. "System card" is the name some companies, such as Anthropic and OpenAI, use for their version, which often goes deeper into safety testing and how the model is released.
Can I trust the numbers on a model card?
They are a useful starting point, but most come from the maker's own tests. Check whether outside testers were involved, and always try the model on a sample of your own work.
Where did the idea come from?
A 2019 research paper by Margaret Mitchell and colleagues proposed model cards as short documents that report how a model performs in different conditions and how it is meant to be used.
Related explainers
The DevDrift helps teams choose AI models and test them on their own work before rollout: see how.
Sources
Could you explain this to a colleague now?
Mark it and posts that mention it will show a tick. Saved in this browser only.