Facebook’s Civil Rights Audit Chapter Six: Algorithmic Bias Algorithms, machine-learning models, and artificial intelligence (collectively “AI”) are models that make connections or identify patterns in data and use that information to make predictions or draw conclusions. AI is often presented as objective, scientific and accurate, but in many cases it is not. Algorithms are created by people who inevitably have biases and assumptions, and those biases can be injected into algorithms through decisions about what data is important or how the algorithm is structured, and by trusting data that reflects past practices, existing or historic inequalities, assumptions, or stereotypes. Algorithms can also drive and exacerbate unnecessary adverse disparities. Oftentimes by repeating past patterns, inequality can be automated, obfuscating and perpetuating inequalities. For example, as one leading tech company learned, algorithms used to screen resumes to identify qualified candidates may only perpetuate existing gender or racial disparities if the data used to train the model on what a qualified candidate looks like is based on who chose to apply in the past and who the employer hired; in the case of Amazon the algorithm “learned” that references to being a woman (e.g., attending an all-female college, or membership in a women’s club) was a reason to downgrade the candidate. Facebook uses AI in myriad ways, such as predicting whether someone will click on an ad or be interested in a Facebook Group, whether content is likely to violate Facebook policy, or whether someone would be interested in an item in Facebook’s News Feed. However, as algorithms become more ubiquitous in our society it becomes increasingly imperative to ensure that they are fair, unbiased, and non-discriminatory, and that they do not merely magnify pre-existing stereotypes or disparities. Facebook’s algorithms have enormous reach. They can impact whether someone will see a piece of news, be shown a job opportunity, or buy a product; they influence what content will be proactively removed from the platform, whose account will be challenged as potentially inauthentic, and which election-related ads one is shown. The algorithms that Facebook uses to flag content as potential hate speech could inadvertently flag posts that condemn hate speech. Algorithms that make it far more likely that someone of one age group, one race or one sex will see something can create significant disparities — with some people being advantaged by being selected to view something on Facebook while others are disadvantaged. When it comes to algorithms, assessing fairness and providing accountability are critical. Because algorithms work behind the scenes, poorly designed, biased, or discriminatory algorithms can silently create disparities that go undetected for a long time unless systems are in place to assess them. The Auditors believe that it is essential that Facebook develop ways to evaluate whether the artificial intelligence models it uses are accurate across different groups and whether they needlessly assign disproportionately negative outcomes to certain groups. A. Responsible AI Overview Given the critical implications of algorithms, machine-learning models, and artificial intelligence for increasing or decreasing bias in technology, Facebook has been building and growing its Responsible Artificial Intelligence capabilities over the last two years. As part of its Responsible AI (RAI) efforts, Facebook has established a multidisciplinary team of ethicists, social and political scientists, policy experts, AI researchers and engineers focused on understanding fairness and inclusion concerns associated with the deployment of AI in Facebook products. The team’s goal is to develop guidelines, tools and processes to help promote fairness and inclusion in AI at Facebook, 76