noun
A model card is a standardized document that describes an AI model’s capabilities, limitations, ethical considerations, and potential biases. It typically includes performance metrics, training data sources, and use-case recommendations.
Developed by researchers and engineers, model cards aim to promote transparency in AI by providing clear, accessible information about a model’s behavior. For example, a language model’s card might disclose that it struggles with sarcasm or has been trained on data up to 2023.
For users, a model card helps assess whether an AI tool is suitable for their needs. It may forbid certain high-risk applications, like medical diagnoses, if the model isn’t validated for them. For developers, it ensures accountability by documenting known flaws upfront. A real-world example: Google’s BERT model card notes its 93% accuracy on the GLUE benchmark but warns about biases in gendered language.
Before using a generative AI tool for legal advice, check its model card to see if it was trained on up-to-date laws.
In other languages
See also: Bias mitigation