The influence of response bias on confidence and accuracy in multi-alternative tasks for humans and artificial neural networks
Bogeng Song, Dobromir Rahnev · Journal of Vision · 2025
How does response bias influence confidence judgments in perceptual decision making? Most previous research has used 2-choice tasks, which makes it difficult to address this question. Indeed, in 2-choice tasks, bias for one category is equivalent to a bias against the other category, meaning that one cannot independently assess the bias for each category. Therefore, to determine how response bias influences confidence judgments, here we used two datasets with multi-alternative tasks. In the first dataset, 60 subjects performed an 8-choice digit categorization task using noisy MNIST images. In the second dataset, 37 subjects performed a 16-choice image classification task using blurred ImageNet images. We quantified the bias for each stimulus category as the rate at which a category was chosen when the stimulus was selected from all other categories. Unlike for 2-choice tasks, this procedure allowed us to estimate the bias for each category independently from the remaining categories. We then used the response bias for each category to predict the confidence and accuracy for that category. We found that a bias towards a given category predicted higher accuracy. Counterintuitively, however, being biased towards a category predicted lower confidence ratings for that category. These results were robust to variations in task difficulty and condition. We further observed that artificial neural networks (ANNs) trained on object recognition replicated the positive effect of response bias on accuracy, but, in contrast to humans, also showed a positive effect of response bias on confidence. These findings suggest that when rating confidence, humans – but not ANNs – actively discount evidence for stimulus categories that they are biased towards. More broadly, our results demonstrate the utility of complex, multi-alternative tasks in discovering the mechanisms driving visual confidence.