5.3 Computing Bias
As we've discussed throughout these guides, computing innovations can reflect existing biases. Biases can be embedded at all levels of development, from the brainstorming phase to the work done after release. This can take the form of a bias written into the algorithm itself or bias in the data used.
For example**, criminal risk assessment tools** are used to determine the chances that a defendant will re-offend, or commit another crime. This information is then used to influence decisions across the judicial process. However, these algorithms are trained to pick out patterns and make decisions based on historical data, and historical data is historically biased against certain races and classes. As a result, risk assessment tools may disproportionally flag certain groups as risks.
Algorithms might also be trained on sets of data that aren't as diverse as they need to be. For example, facial recognition systems are often trained on data sets that contain fewer images of women and minorities than men in the majority.
Finally, computing innovations use data from the world around them, and that world is often biased in its own right.
People can take steps to combat these biases. They can be mindful of the potential for bias and make sure that their data is as unbiased and representative as possible. This is not only good for the program itself, but also for society as a whole. After all, algorithms are written by people, and being able to find and eliminate bias in computers can help us eliminate bias in ourselves as well.