Model Cards to Support Responsible AI

A model card is a brief user guide to any algorithms or data processes that will be used by colleagues in your organisation. It is a way for the data team to communicate what they have built, share insights into best practice and include warnings about any limitations and caveats around future uses of the data process. Clear documentation is an essential part of Responsible AI and model cards can help support transparency and communication.

It is up to the data team to determine which model card or fact sheet best suits their ways of working, but we have created a sample model card to guide our members through documentation best practices.

You can access the blank model card here. Use this form to start taking notes to use for your final model card.

You can view a sample of a completed model card here.

Tips for effective model cards

• • Assume that the audience will be the end user eg colleagues in marketing or customer services or possibly even clients.

• • Assume that the intended use and consequences will be obvious to most end users. Your job is to highlight the concerns you would have around unintended uses and consequences which can be less visible to those not familiar with the affordances and limitations of working with data.

• • Do not avoid technical terms where they are necessary to communicate clearly. Offer a brief explanation of terms such as precision or recall and their significance, where appropriate. It is best to share this essential information and if comprehension is a problem, discuss ways to address gaps in staff training around data use with your human resources team.

• • Include contact information for people who would like to learn more about the model or who have concerns about the way that the model is being used.

• • Keep any descriptions brief. You may be very proud of your work to incorporate non-normal distributions of the data or the cleaning and imputation processes that you used. Unless these are likely to impact on the final use of the algorithm, this is not the place or the audience to share those insights with. Just focus on how the end user should and shouldn’t work with the algorithm.

• • Ask colleagues to review it. It is the nature of unintended consequences that they are very difficult to spot in advance. A team will perform better at anticipating all possible misuses.

For more on data best practices, see our Governance & Professionalism training which is provided as a benefit of IoA Membership.