How Do You Test the IQ of an AI—or a Superintelligence (SI)?

What if an AI can produce a brilliant analysis—and still misunderstand what it is analyzing? An AI system may reason convincingly, apply sophisticated evaluation criteria, and deliver a seemingly authoritative judgment—while entirely missing the actual subject of its assessment. The problem may not be insufficient information or weak reasoning. It may be the unrecognized substitution of an unfamiliar subject with a familiar but inadequate interpretive model. A wrong model does not become right merely because an AI has mastered it. This conceptual paper introduces Qualitative Intelligence in the sense of Knowledge Quality (KQ): the ability to understand unfamiliar subjects on their own terms, recognize one's own interpretive and evaluative models as fallible representations, and autonomously revise an inadequate framework when its limitations become evident. It challenges a fundamental assumption behind many intelligence evaluations: Before measuring how successfully an AI solves a problem, shouldn't we determine whether it has correctly understood what the problem—or the subject under investigation—actually is? A New Perspective on Measuring AI Intelligence The paper proposes a general, source-grounded First-Contact Test. An AI system receives a complete, challenging source without substantive interpretive guidance. Its initial analysis is then evaluated against an independently established, source-grounded reference. The approach distinguishes genuine reconstruction of the subject from logical evaluation and legitimate external criticism. It does not reward agreement with an author; evidence-based disagreement remains entirely permissible. Three progressively extended implementations are proposed: Core: Measures the quality of the autonomous initial interpretation against a predefined source reference. Core+: Examines improvement after exactly one content-neutral self-review opportunity. Full: Introduces controlled conditions to investigate framework dependence, the loss of valid alternatives through Qualitative Inhibition (QI), and possible self-stabilization of flawed interpretations through the Qualitative Prisoner's Dilemma (QPD). The first concrete reference case is the complete 77-page, three-part foundational Knowledge Quality publication by Dr. Thomas R. Glück. The general testing method, however, is not restricted to KQ or to any particular discipline. The paper also outlines a possible AI-assisted evaluation architecture, in which a separately validated AI scoring system applies an independently established reference and coding key to another AI's initial response. Why This Matters As AI systems become increasingly capable—and as the prospect of superintelligence (SI) raises the stakes of evaluation—performance on familiar tasks may be insufficient to establish whether a system has understood the subject it is reasoning about. The proposed framework therefore opens a complementary research direction at the intersection of artificial intelligence evaluation, intelligence measurement, epistemology, and the quality of conceptual representation. Scientific Status and Research Invitation This is an original conceptual research proposal, not an empirically validated or population-normed IQ test. The historical AI interaction discussed in the paper is a retrospective illustration, not an independently validated measurement or causal diagnosis. No superintelligent system is claimed to have been tested. The proposed methodology invites independent replication, blinded reference construction, controlled experiments, and transfer studies using unfamiliar sources beyond KQ. The central question is not simply how intelligent an AI appears to be within its model. It is whether the AI can recognize when the model itself is wrong.