Skip to main content

Issue #1 | Summer 2022

Model Cards

Model cards are used to document Machine Learning models. They are intended to provide information about the context in which algorithmic decision systems are used. The short documents detail the performance characteristics of the respective algorithm in a structured manner and also contain information about the context in which model training took place, including information about different cultural, demographic or phenotypic groups (ethnicity, geographic locations, gender or Fitzpatrick skin type, for example) and intersectional groups (such as age and ethnicity or gender and Fitzpatrick skin type). Complete documentation should also include the type and details of the Machine Learning model as well as the intended use and possible influencing factors. In addition, test and training data should be recorded on model cards along with any ethical issues or concerns. The aim of the documentation is to ensure that the Machine Learning behind the models in question becomes more transparent and comprehensible. Google, for example, has published a model card for an algorithm that recognizes faces in photos and videos. Another example is the model card for the commonly used language processing model BERT, published on the developer platform Hugging Face.