This paper defines Cognitive Drift and Co-Cognition as a framework for understanding how AI-mediated reasoning changes human thought. It focuses on what happens when cognition becomes distributed across people, prompts, and outputs. The paper gives language for patterns like AI dependency, externalized memory, recursive thinking, and the feeling that thinking with AI is different from simply using a tool.
Modern systems increasingly remain operational while producing outcomes that feel hollow or disconnected from reality. This paper proposes Reality Drift: a structural failure mode where representations, metrics, and models drift away from reality faster than corrective constraints can bind them. The result is systems that continue functioning while losing the ability to self-correct.
Large language models can keep producing confident, well-structured answers even when they’re wrong. Because mistakes don’t slow them down or force revision, errors tend to accumulate quietly instead of triggering correction. This piece explains why that happens and why fluency alone is a weak indicator of alignment.