The recent AI-based Wigner's friend experiment proposed by Zeng and Zhang (2025) attempts to test the Absoluteness of Observed Events (AOE) using a deep learning model as a human-like observer. However, their framework introduces a potential logical circularity by coupling the verification of AOE with the qualification of the AI as a legitimate observer. We propose a simplified gedankenexperiment that circumvents this issue entirely—not by validating the AI as an observer, but by rendering such validation unnecessary for testing AOE. By pre-training an AI with an internal law derived exclusively from deliberately fabricated unentangled measurement statistics, and equipping it with a pre-reflective hard filtering mechanism, we create a scenario where the final observed data alone can distinguish between two worldviews. If the data perfectly conforms to Born rule statistics, it supports the failure of AOE (observer-dependent reality). If the data exhibits macroscopic structural anomalies, it supports AOE (observer-independent reality). This scheme eliminates the need to independently establish the AI's observer status, providing a cleaner logical probe into the foundations of quantum reality.
Shuhui Zhang (Fri,) studied this question.