INSUFFICIENT DATA FOR A MEANINGFUL ANSWER Epistemic Humility and Temporal Deferral in Asimov's Multivac as a Proto-Ethics of AI Systems

This conceptual essay explores hallucinations in large language models (LLMs) not as a purely technical flaw, but as the predictable outcome of how we train, evaluate, and culturally frame “intelligent” systems. Through three thought experiments on logical and mathematical knowledge, the paper shows how different training regimes (true-only, true/false, and true/false/unknown) reshape not just accuracy, but the forms of error and bluffing. The second part reads Isaac Asimov’s “The Last Question” as a proto-ethics of AI: Multivac’s famous reply, “INSUFFICIENT DATA FOR A MEANINGFUL ANSWER”, is proposed as a model of epistemic humility and temporal deferral for future AI systems.