False Epistemic Redundancy: Do AI Ensembles Share a Blind Spot?

When multiple AI personas that differ in terms of cognitive reasoning appear to explore a problem differently, how much genuine epistemic coverage is actually gained and which tails might they still collectively miss? We explore whether prompting the same AI from different perspectives broadens the hypotheses it considers, particularly toward the less obvious “tails” where rare but consequential explanations may lie. Across our real-world cases, no persona consistently recovered the true explanation. Instead, success varied by case and framing, even when agents were given the evidence that ultimately led human investigators to the solution. At the same time, persona prompting on the basis of causal, analogical, teleological, and dialectical reasoning consistently increased the apparent diversity of responses. In this small pilot, greater response diversity therefore did not reliably correspond to better coverage of the consequential hypothesis, implying that surface level model prompting may be insufficient at inducing epistemic diversity, calling for methods that target deeper model activations.