Preprints

Large language models, like humans, struggle to overcome the “curse of knowledge” and ignore potentially biasing information. However, unlike humans, they possess the ability to “self-blind” by calling their own API with an appropriately redacted prompt. In this collaboration with Brian we show that, when given the opportunity, they do just that.

In recent years, more and more mental health research relies on large-scale online testing. Researchers get hundreds of participants to fill out many mental health questionnaire, and then identify links between their psychiatric tendencies (as revealed by their questionnaire responses) and their behaviour in controlled tasks. Using this method, previous research identified robust links between mental health (compulsivity, depression and anxiety) and decision confidence. In this collaboration with Noam and Ruvi, we show that these correlations are largely driven by factors outside mental health, namely participants’ response biases and their level of attention.