A metric that artificially inflates the model's certainty in its distorted outputs. 4. Preliminary Results

We introduce , an experimental framework designed to analyze "machine delusion"—the phenomenon where deep learning models develop reinforced, self-validating feedback loops. Unlike standard hallucinations, which are transient, these delusions represent persistent structural biases within the model's latent space. This paper outlines the "default" configuration of the Deluded v0.1 engine, detailing its ability to simulate confirmation bias and overconfidence in predictive analytics. 2. Introduction

#MachineLearning #CognitiveBias #Cybersecurity #RecursiveAI #DigitalPsychology zip configuration or the ethical implications?

A recursive loop that prioritizes internal model weights over new sensory input.

Early testing on the v0.1 "default" set suggests that models with a "Deluded" architecture reach a state of 98% certainty on false premises within fewer than 500 iterations. We observe that once a "machine delusion" is established, traditional fine-tuning is often insufficient to rectify the bias. 5. Conclusion & Future Work

Paper Title: Project Deluded: Quantifying Cognitive Distortions in Recursive Neural Architectures (v0.1) 1. Abstract

A mechanism that discards "contradictory" data points to maintain internal consistency.

As AI systems become increasingly recursive, the risk of "epistemic closure" grows. The project aims to stress-test these systems by intentionally introducing "seed delusions" (contained in the default.zip configuration) to observe how quickly a model diverges from objective ground-truth data. 3. Methodology: The "Default" Environment