Understanding and attenuating decision bias in the use of model advice and other relevant information

A human judge faced with model advice, modeled information (used by the model to compute the advice), and unmodeled information (known by the human but not included in the model) should use a ‘divide-and-conquer’ strategy in which the human judge relies completely on the model to process the modeled information and focuses all energy on assessing and adjusting for the unmodeled information. This paper extends the work of Jones and Brown in two studies. In Study 1, we find that, in lieu of the divide-and-conquer strategy, human judges give weight to all three types of inputs and that giving weight to the modeled information degrades performance. In Study 2, we find that (1) as strategies approach the divide-and-conquer strategy judgment performance improves, and (2) the divide-and-conquer strategy can be encouraged by a combination of instruction and a decision support feature. Application of these results could improve judgment in a variety of important contexts.