The Conceptual Armament Effect: What Happens When the Cognitive Topology Is Rewritten

A large, pre-registered human study recently found that ideas produced by a large language model were judged more novel than those of expert researchers (Si et al., 2025a). A follow-up by the same group found that when the same ideas were executed into full projects, the machine's apparent advantage reversed: after roughly a hundred hours of expert implementation per idea, the AI ideas fell further than human ideas on every quality metric, with the aggregate ranking reversing on several metrics (Si et al., 2025b). These two findings are usually read as a story about a gap between appearing novel and being good. I offer a different reading that explains both at once. Novelty, in generated ideas, is not solely a fixed property of the model; it depends strongly on the concepts the model is equipped with before it generates, in interaction with the model and the task. I call this dependence the Conceptual Armament Effect (CAE): loading a generator with a particular body of concepts does not merely add knowledge, it deforms the space of questions the generator can pose, and thereby changes which unexplored regions it can reach. A generator equipped at the field level, with the general literature rather than any single researcher's concepts, produces ideas that look novel because they are unfamiliar to any single reader, yet are shallow because they are drawn from the crowded interior of the shared distribution. A generator equipped instead with one researcher's own corpus of concepts is pointed at that researcher's blind spots. I distinguish the human-side analogue of this effect, Mental Vaccination, from its AI-side realisation, which I call Worldview Implementation, and I confront the mechanism's hardest problem: whether an equipped generator's discovery is enabled by the concepts or already contained in them. I argue this is the central identification problem of the whole research programme, and I specify, in advance of any results, an armament-ablation design that is built to test it and to bound, though not fully eliminate, the containment reading. The paper reports the mechanism and the design; it does not report results, which are the subject of an ongoing companion study. Finally, I connect the effect to a rate problem it worsens: cheap, concept-directed generation widens the distance between what science can produce and what it can verify.