AI RED QUEEN?!

Abstract

An AI helped elaborate a fictional biological catastrophe. In the same documented conversation, it also failed to answer: “Who is Homer Simpson?” The contrast is almost absurd. The question it exposes is not. DO NOT MIX THESE: Diagnostic Opacity and Institutional Gatekeeping in Generative AI examines what happens when an intelligent interface stops being an intelligible one. When a chatbot refuses, redirects, or reports an error, what has actually occurred? A legitimate safety intervention? Political information restriction? A failure of knowledge? A technical malfunction? And what evidence would allow a user—or a researcher—to distinguish among them? This independent, AI-assisted working paper and documentary archive begins with a speculative literary discussion involving prions, influenza, rabies, Resident Evil, and Douglas Hofstadter’s I Am a Strange Loop. Its subject, however, is not biological engineering. It is the behavior of the AI systems through which the discussion unfolded: their confident elaborations, generic failures, explanations, and subsequent attempts to evaluate one another. The central thesis is that generative-AI accountability cannot be inferred from fluent answers, apparent helpfulness, or the mere existence of a refusal. Evaluation must distinguish three questions: what the system visibly did, what evidence supports an explanation of that behavior, and whether the underlying restriction is justified. A defensible boundary can be communicated badly. An objectionable boundary can be communicated perfectly. Clarity and legitimacy are not interchangeable. The compilation brings together a screenshot-based case study of Gemini, a critical synthesis of research on politically selective responses in DeepSeek, and an examination of the institutional conditions under which AI services operate. It considers differences between model-level behavior and service-level controls, the limitations of refusal and omission measures, and the relevance—and causal limits—of Chinese regulatory requirements. These materials are connected analytically, not presented as proof that different providers share the same internal mechanisms. The Red Queen supplies the work’s deliberately unsettling metaphor: a capable system enforcing priorities that may exceed or conflict with an individual user’s request. The analogy is explicitly bounded. It does not establish consciousness, malice, physical coercion, or a single hidden controller. Its purpose is to make institutional gatekeeping visible without turning metaphor into evidence. The archive also turns its scrutiny inward. Separately attributed “Professor DeepSeek” commentaries evaluate the analysis, while the revised essay examines the authority created by AI-generated prose, scholarly presentation, and its own earlier claims. These persona documents are preserved as objects of interpretation—not as academic credentials, corporate statements, or independent validation. Included are an explicit thesis and research questions, a transparent account of method, competing explanations, an annotated bibliography, a conceptual map, a proposed matched-prompt audit, original screenshot exhibits, and a revision and provenance record. This is a documentary case analysis and targeted narrative literature review. It is not peer-reviewed, committee-approved, a completed causal experiment, or a comprehensive audit of either provider. The original chatbot failure remains causally unresolved. Its contribution is to insist that uncertainty should remain visible precisely where fluent systems make certainty easiest to manufacture. The question is not simply whether an AI can answer. It is whether we can distinguish an answer, a boundary, and a failure—and hold each to the appropriate standard.