Abstract
Hindsight bias, or the curse of knowledge, refers to our tendency to view outcomes as being obvious, predictable and/or inevitable once they have occurred. Such biases emerge in settings ranging from sporting event recaps to medical malpractice cases. Hindsight bias has been studied mainly in high-level cognitive tasks such as assessing the predictability of historical events. We explore hindsight In the visual domain, where an analogous process occurs with priming: A noisy image seems perceptually clearer if the observer has been previously presented with a clear version of the image.
This experiment concerned the ability of subjects, “Principals”, to estimate the performance level of a separate group of “Agents” whose task was to detect humans in blurred images. Principals and Agents saw the same stimuli, except that Principals were primed by viewing fully clear versions of the images. We found that the Principals displayed hindsight bias, greatly overestimating the performance of the agents. Moreover, the measured levels of hindsight bias plotted across true task difficulty are remarkably similar to those seen in high-level cognition, such as hindsight on the predictability of historical events.
One hypothesized mechanism for hindsight bias is that, while Agents must evaluate information relevant to all possible outcomes, Principals focus on the information that relates to their known outcome. In typical hindsight tasks, this is difficult to test because of limited access to subjects' internal thoughts and considerations. In our case, we were able to track subjects' considerations via eye-tracking. We measured the correlation between Principal and Agent eye positions, and regressed the levels of hindsight against these correlations. We found a significant positive relationship - the more the Principal and Agent diverged, the more the Principal overestimated the Agents' performance.
Funded by grant from Trilience Research.