Facebook’s Civil Rights Audit and make these resources widely available across the entire company so there is greater consistency in approaching questions of AI fairness. During the Audit process, the Auditors were told about Facebook’s four-pronged approach to fairness and inclusion in AI at Facebook: (1) creating guidelines and tools to identify and mitigate unintentional biases; (2) piloting a fairness consultation process; (3) participating in external engagement; and (4) investing in diversity of the Facebook AI team. Facebook’s approach is described in more detail below, along with and observations from the Auditors. 1. Creating guidelines and tools to identify and mitigate unintentional biases that can arise when the AI is built and deployed. There are a number of ways that bias can unintentionally appear in the predictions an AI model makes. One source of bias can be the underlying data used in building and training the algorithm; because algorithms are models for making predictions, part of developing an algorithm involves training it to accurately predict the outcome at issue, which requires running large data sets through the algorithm and making adjustments. If the data used to train a model is not sufficiently inclusive or reflects biased or discriminatory patterns, the model could be less accurate or effective for groups not sufficiently represented in the data, or could merely repeat stereotypes rather than make accurate predictions. Another source of potential bias are the decisions made and/or assumptions built in to how the algorithm is designed. To raise awareness and help avoid these pitfalls, Facebook has developed and continues to refine guidelines as well as a technical toolkit they call the Fairness Flow. The Fairness Flow is a tool that Facebook teams use to assess one common type of algorithm. It does so in two ways: (1) it helps to flag potential gaps, skews, or unintended problems with the data the algorithm is trained on and/or instructions the algorithm is given; and (2) it helps to identify undesired or unintended differences in how accurate the model’s predictions are for different groups or subgroups and whether the algorithms settings (e.g., margins of error) are in the right place. The guidelines Facebook has developed include guidance used in applying the Fairness Flow. The Fairness Flow and its accompanying guidelines are new processes and resources that Facebook has just begun to pilot. Use of the Fairness Flow and guidelines is voluntary, and they are not available to all teams. While the Fairness Flow has been in development longer than the guidelines, both are still works in progress and have only been applied a limited number of times. That said, Facebook hopes to expand the pilot and extend the tools to more teams in the coming months. Facebook identified the following examples of how the guidelines and Fairness Flow have been initially used: • When Facebook initially built a camera for its Portal product that automatically focuses the camera around people in the frame, it realized the tracking did not work as well for certain genders and skin tones. In response, Facebook relied on its guidelines to build representative test datasets across different skin tones and genders. Facebook then used those data sets on the algorithm guiding the camera technology to improve Portal’s effectiveness across genders and skin tones. 77