AI Policy, Law & Safety · AI Regulation
What is a model card and is publishing one legally required anywhere
A model card is a standardized document describing an AI model's intended use, known limitations, and training data characteristics, and while widely adopted as a voluntary industry best practice, some emerging regulations, including aspects of the EU AI Act, have begun requiring comparable documentation for certain higher-risk AI systems.
Key takeaways
- A model card documents an AI model's intended use, known limitations, and training characteristics.
- Publishing model cards began as a voluntary industry best practice, not a universal legal requirement.
- Some emerging regulations now require comparable documentation for certain higher-risk AI systems.
- Requirements and enforcement vary considerably depending on jurisdiction and the system's risk category.
What a Model Card Actually Documents
A model card is a standardized document, typically published alongside a released AI model, describing its intended use cases, known limitations and failure modes, and relevant characteristics of the data it was trained on — giving developers and users a clearer picture of what a model is actually suited for before they deploy it.
How This Started as a Voluntary Practice
Model cards began as a voluntary industry best practice, championed by AI researchers and responsible AI advocates as a way to improve transparency around model capabilities and limitations, without any legal requirement forcing companies to publish them in this specific standardized format.
Where Regulation Has Begun Requiring Similar Documentation
More recently, emerging AI regulations, including aspects of the European Union’s AI Act, have begun requiring comparable documentation for certain higher-risk AI system categories, effectively converting what was previously a voluntary best practice into a legal obligation for systems falling within specific regulated risk tiers.
Why Practice Still Varies Considerably
Despite this regulatory movement, actual practice still varies considerably across companies and model types, since many models — particularly those not falling into a specifically regulated high-risk category — still aren’t subject to any legal documentation requirement, leaving publication largely at each company’s own discretion.
Why This Documentation Genuinely Matters
Regardless of whether it’s legally required in a given context, thorough model card documentation genuinely helps downstream developers and users make more informed decisions about whether a specific model is actually well-suited to their intended use case, reducing the risk of deploying a model in a context it wasn’t designed or tested for.
Bottom Line
A model card documents an AI model’s intended use, limitations, and training characteristics, and while it began purely as a voluntary industry best practice, some emerging regulations like the EU AI Act have started requiring comparable documentation for certain higher-risk systems specifically.
Go deeper
Frequently asked questions
Do all AI companies publish model cards for every model they release?
No — practice varies considerably across companies and model types, with some organizations publishing detailed model cards consistently while others provide much less detailed documentation, particularly for models not subject to specific regulatory documentation requirements.
Related questions
- What Is the NIST AI Risk Management Framework?
- What Is the EU AI Act and Who Does It Apply To?
- What Is a 'High-Risk' AI System Under EU Regulation?
- What is a compute threshold and why do some ai regulations use it to determine oversight?
- How do different countries define what counts as a high risk ai system?
- Can you be held liable for relying on incorrect advice from an ai tool?
Sources
- [1]AI standards and risk framework research — National Institute of Standards and Technology
- [2]European digital policy and regulation — European Commission
Written by Editorial Team
Last updated July 30, 2026
Get one well-sourced answer a week
No spam. Unsubscribe anytime.