What is a model card, and why am I hearing about them?

A model card is the nutrition label for an AI tool

A model card is a short document the company behind an AI tool writes about it. It says what the tool is for, what it was trained and tested on, how well it performs and for whom, where it fails, and who is responsible for it.

You will never write one. You may be the person who needs to read one, because you are the one putting a client’s or patient’s information into the tool.

Why you keep hearing about them

Model cards have moved from a research idea to a document regulators, states and vendors are arguing over. Four things put them in the news:

• Where they came from. Researchers proposed model cards in 2018 as a standard way to report what an AI model does and where it falls short (Mitchell et al.).
• A federal rule required them in EHR software. Since a 2023 federal rule, certified electronic health record software has had to disclose details about its AI decision-support features to the clinicians using them.
• That rule is now gone. This happened on October 26, 2026. The federal rule called HTI-5 was finalized that day, and it removes the AI model card requirement from the certification program (HHS, AHA). The agency had proposed removing 34 certification criteria and finalized 14. It kept the privacy and security criteria after comments came in. It did not keep the model card. California’s Attorney General had formally objected in February 2026 (letter). So there is no federal disclosure requirement left, and whatever a vendor chooses to publish is what you get.
• Industry groups are writing voluntary versions. Coalitions of health systems and technology companies now publish their own templates. One example is the Coalition for Health AI, whose template and registry are public. Voluntary means the vendor decides what goes in.

What this means for you: the model card is becoming the document everyone asks a vendor for. Health systems have staff to read them. You have this page.

What it means for a practice of one

A model card is the document to ask for before client information goes into a tool. That includes note-takers, scribes, chatbots, scheduling assistants, and AI features built into software you already pay for.

You do not need to understand every line. You need to know whether the tool was built and tested for work like yours, and what the company says it should not be used for. If your use sits outside what the card describes, the risk is yours, not the vendor’s.

How to: read one in five minutes

Skip the technical sections. Look for these five things:

1. Intended use. What the company says the tool is for. Is your use on that list?
2. Out-of-scope use. What it says the tool should not be used for. This is often the most useful line on the card.
3. Who it was tested on. Ages, conditions, settings. A tool tested in hospitals may not behave the same in a coaching or nutrition practice.
4. Known limits. Where it makes mistakes, and what kind.
5. Data handling and contact. What happens to the information you enter, and who to contact when something goes wrong.

Write down what you find, with the date. That note is your record that you checked.

How to: get answers when there is no model card

Most small vendors do not publish a model card yet. That is not a reason to stop using the tool. It is a reason to ask. Send these three questions by email, so the answers are in writing:

1. What is this tool designed to do, and what should it not be used for?
2. What happens to the information I enter: is it stored, and is it used to train your AI?
3. Who do I contact if the tool gives a wrong or harmful answer?

A vendor that cannot answer these three in plain words has told you something too.