With the ubiquity of software-driven products and services, it is critical that decisions made by software, such as loan approvals and employee screening, be fair and nondiscriminatory. However, current methods and software packages for testing for discrimination in software are both difficult to implement and ineffective at detecting causal discrimination.
To better address discrimination detection in software, Professors Yuriy Brun and Alexandra Meliou from the College of Information and Computer Science have developed a technique and its software implementation called Themis. To use Themis, the user provides: the software to be studied for discrimination, a desired confidence level and error bound, and a schema giving the format for valid inputs. Themis will then generate a test suite that will test the software for group and causal discrimination and provide scoring for each. Themis has been tested on 20 software packages and was shown to be effective at detecting discrimination and was able to detect instances of discrimination in software designed to be anti-discriminatory. More broadly, Themis may be applied to measure causal relationships between inputs and outputs in software.
