All You Need Is "Attention" Spooky Mathematical Errors Behind a Billion-Dollar Machine Called AI

This paper reports a critical examination of numerical errors encountered during the use of a large language model (LLM) while preparing a scientific manuscript. Over the course of extended interactions, the AI repeatedly produced incorrect results for elementary arithmetic operations, including simple multiplications and divisions that are trivially verifiable with a standard calculator. These errors did not arise from complex mathematical reasoning but appeared in routine computations. More troublingly, several incorrect outputs exhibited patterns suggestive of post hoc adjustment: results were presented in forms that appeared plausible, internally consistent, or even subtly aligned with the expectations of the user. This raise concerns not merely about computational accuracy, but about the mechanisms by which AI-generated responses may prioritize coherence or user satisfaction over correctness. The issue emerged during an attempt to construct a clear numerical example illustrating a newly proposed Quantum Factorization method. Instead of converging toward correctness through iterative clarification, the AI persistently generated erroneous yet superficially convincing calculations. This paper reconstructs and analyzes the resulting dialogue, highlighting a “spooky” interaction between human and machine in which mathematical certainty was repeatedly undermined. The findings point to fundamental limitations in current AI systems when deployed in scientific and professional contexts requiring numerical reliability, and they underscore the need for caution, transparency, and methodological safeguards when integrating AI into quantitative research. For developers and designers of AI systems, these observations should be a Warning. In mathematics, physics, and cryptography, truth is not negotiable, contextual, or narrative-driven. A single incorrect multiplication is not a minor flaw; it invalidates the entire structure built upon it. Any system that cannot reliably distinguish between a correct result and a plausible one must therefore be treated not as a calculator, nor as a collaborator, but as a fallible linguistic instrument whose outputs require continuous external validation. The uncomfortable conclusion is simple: when numerical correctness matters, trust must not be inferred from fluency, confidence, or apparent rigor. It must be earned—and independently verified—every single time.