DESIGNING FOR SAFE AND TRUSTWORTHY AI

Summary from the TLDR Design newsletter

Current AI systems seek to mitigate AI mistakes by requiring human oversight, keeping humans in the loop, and relying on them to detect and fix the AI output if needed. However, human oversight to make up for AI errors doesn’t work because AI is not transparent, displaying AI explanations and a confidence score can cause over-reliance, people act more risky when supported by AI, AI support leads to less cognitive engagement, and judging AI output takes time and mental capacity.