Abstract
Information integration for perception and decision making can be near-optimal, such as reliability-weighted averaging of sensory cues (Ernst and Banks, 2002) or sensory cues and prior knowledge (Wolpert et al. 2011). Behaviour diverges from optimal in a variety of circumstances (Rahnev and Denison, 2018). Recent studies suggest suboptimalities arise as perceptual and decision-making systems are not equally sensitive to all sources of uncertainty. Castanon et al. (2018) suggest un-certainty brought about through sensory encoding noise can be accounted for, but uncertainty in higher level integration processes cannot. Similarly, Kiryakova et al. (2020) hypothesise that uncer-tainty arising within perceptual/decision-making systems (intrinsic noise) can be tracked, but un-certainty arising in the world (extrinsic noise) cannot. Here, 40 participants used intrinsic-only or intrinsic+extrinsic noise visual cues (dot-clouds) and prior information (Gaussian base-rate distribu-tions shown and reinforced through feedback) to estimate the location of a hidden target. Intrin-sic-only cues were four dots from a Gaussian centred on the true location with low/high variability (low/high intrinsic noise). Intrinsic+extrinsic cues were centred on a position varying about the true location according to a draw from a second Gaussian, adding extrinsic uncertainty. Participants bi-ased their responses more towards the prior when presented with high compared to low intrinsic noise cues (p < .001) but did not adjust the weight given to the cues when extrinsic noise was add-ed (p = .209). Accordingly, the weight placed on the cue was further from optimal when extrinsic noise was present (p < .001). Subjective uncertainty measures suggest participants were not fully aware of the extent of the added extrinsic noise. These results are in favour of the hypothesis that perceptual and decision-making systems struggle to track and account for extrinsic noise. They suggest optimal information integration for perception and decision-making is only possible when accounting for specific types of uncertainty.