Testing the effectiveness of using psychology-intuitive information to promote responsible AI: information about the training set shapes perceptions of AI algorithms in an experimental setting

Most artificial intelligence (AI) stakeholder groups, including administrations, institutions and expert groups, agree on the need to ensure the responsible use of AI. Of the available tools that can be used to promote responsible AI, most regulations target AI designers, adopting a top-down approach to responsible AI and overlooking the role of end-user perceptions. However, emerging evidence shows that end users have preferences for specific types of algorithms, suggesting that their perceptions may be leveraged to promote responsible AI. Nevertheless, aligning end users' intuitions with the actual performance — particularly the reliability — of algorithms remains challenging given the number and complexity of the features impacting AI algorithms' reliability. In this paper, we propose to build on decades of research in psychology to identify which information that is important from a technical perspective can also be used intuitively by individuals. Using this approach, we demonstrate through a set of five studies involving over 1,000 participants that providing information about the quantity and quality of an algorithm's training set can align end users' perceptions with the actual reliability of AI, thereby offering a tool to promote responsible AI.