A two barplots side by side.The left plot shows accuracy of information retrieval (y-axis, ranging from 0 to 100%) across three conditions (No AI, AI, and AI providing wrong answers). The right plot shows self-reported confidence in the answer correctness (y-axis, ranging from 0 to 5) across the same three conditions. Across the two plots, each condition contains two bars: one for users who watched the relevant segment of the video containing the answer to a question, and one for users who did NOT watch the segment. The plot conveys the following messages. 1) In the "No AI" condition, people who did not watch the video provide far less accurate responses. 2) In the "AI" condition, accuracy is high regardless of whether people watched the video or not. 3) in the "AI providing wrong answer" condition, accuracy drops to the minimum values for people who did not check the video. 4) Self-reported confidence in the answers is always high across all conditions, with just a tiny difference between the watched vs. non-watched scenarios
AI can now answer questions about video content, transforming how people consume information on the Web. An experiment with 900+ participants found that AI improves speed and accuracy of retrieval BUT, users become overreliant on AI answers, accepting wrong answers without any loss of confidence.
