Designing Transparent AI

Although AI offers a promising opportunity to improve many areas of our lives, an issue with the AI-based systems is that these are less than 100% reliable (e.g., Yampolskiy & Spellchecker, 2016). Systems with less than perfect reliability still have the potential to provide significant benefit to users (e.g., Wickens & Dixon, 2007), but this is dependent on the user’s ability to understand the system’s processes and limitations. Further, providing information to users about the reliability of the AI helps calibrate the appropriate level of user trust (e.g., Lee & See, 2004). Unfortunately, many of today’s AI systems are essentially “black boxes” that are difficult to interpret, understand, and trust (Steinruecken et al., 2018). Designers should therefore strive to make various elements of system performance transparent to users. In this talk, we will discuss and provide examples of the following techniques that can be adopted to make AI algorithms and vulnerabilities more transparent to users: