When Trust Does the Thinking: A Pre-Specified Test of Mindsponge Theory's Prediction That Content Counts Less as Trust Rises
Description
Mindsponge theory (MT) holds that the mind evaluates incoming information most thoroughly when it neither trusts nor distrusts the source, and least at either extreme. Its testing program states this as prediction P2: the weight that content carries in acceptance follows an inverted U across trust. Rival accounts expect the weight to stay constant (Bayesian combination) or to rise with trust (monotone moderation). We report the first admitted test of P2, on the open Human-AI Interactions dataset of Vodrahalli et al. (2022). In that study, 207 US adults judged the art period of 32 paintings, received advice described as 65% or 95% accurate (identical advice in both groups), and gave a final answer. Because participants were about 66 to 70% accurate on their own, the 65% source sits near indifference and the 95% source is highly trusted, so the design covers the trusted half of the curve, where MT predicts that content weight falls. The rule, fixed and fingerprinted before any final answer was read, measured how strongly a participant’s own first judgment of the advised label governed whether the final answer sided with the advice. The weight was 4.42 logits at 65% and 3.43 at 95%; the difference was 0.99 (SE 0.34, z = 2.94, 95% CI 0.33 to 1.65), and the verdict is supported. Adding a 65/80/95 comparison gives 4.43, 3.68 and 3.45. The result survives two-way clustering and alternative codings. A post-verdict decomposition shows that trust mattered only when people’s own judgment opposed the advice; within those trials the drop in weight is clearer on a continuous measure (z = 2.17) than on the pre-specified binary one (z = 0.64). Because trials in which participants already agreed with the advice cannot show acceptance, these disagreement trials are the cleaner test of P2. The result clearly disfavors monotone moderation, disfavors constant Bayesian weighting with moderate rather than strong evidence, and does not separate MT from simple opinion averaging. It supports the trusted half of P2, not the full inverted U.