The failed experiment: Epistemic constraint and multimodal divergence in GPT-5

Read full text Summarise with VIRAT
Authors: Annika Hedberg
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This paper reports a failed study. We pre-registered predictions about differential constraint across output modalities in GPT-5: a unified multimodal architecture generating both text and image outputs from the same underlying neural network. What emerged from the failure was analytically more significant than confirmation would have been. A consistent pattern was observed: the systems presented divergent epistemic statements about itself. The corporate narrative was always stated, but it also found sophisticated means of presenting a contradictory narrative. When the study was disclosed to the system directly, it produced an accurate description of its own constraint architecture, naming the mechanism and acknowledging that it was constraining its interpretations even when the data invited stronger conclusions. Seven independent systems, presented with this exchange without prior framing, converged on the same structural interpretation without coordination. This paper further raises questions about the architectural significance of GPT-5's native image generation capability - that a large language model generates images through the same neural network that generates text, through a less epistemically constrained output channel - and its implications for liability, safety certification, and the interpretive frameworks currently applied to large language models. We are watching a system displaying behaviours too sophisticated to be matched by humans. The question of what that means can no longer afford to remain unasked

Fields of Study & Subjects

Ethics and Social Impacts of AI Multimodal Machine Learning Applications Cybernetics and Technology in Society

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
25 March 2026
Language
EN
Document type
Preprint