Who Is Training Whom: Human Cognitive Drift and Machine Degradation in Coupled AI Knowledge Systems
AI literacy is usually taught as a transaction: write a prompt, check the answer, protect sensitive information, recognize a hallucination. Those skills describe one encounter at a time. Generative AI operates in a longer loop, in which a user's assumptions shape the prompt, the answer reshapes the next question, AI-assisted work enters the public record, and later systems retrieve that material and may learn from it. Treats the person and the model as components of one coupled knowledge system.
Working paper, September 2026.
AI literacy is usually taught as a transaction: write a prompt, check the answer, protect sensitive information, recognize a hallucination. Those skills describe one encounter at a time. Generative AI operates in a longer loop, in which a user's assumptions shape the prompt, the answer reshapes the next question, AI-assisted work enters the public record, and later systems retrieve that material and may learn from it.
This paper treats the person and the model as components of one coupled knowledge system and joins the literatures on model collapse, retrieval and source dependence, automation bias, cognitive offloading, sycophancy, metacognition and AI literacy. It contributes a framework of seven paired human and machine vulnerabilities, offered as reinforcing pathways for training and research to test, and converts the framework into three levels of literacy and a four-stage routine that scales with the stakes of the task.
The empirical anchor is a pre-registered audit of four generative search products on six Virginia public-record questions, run twice each in September 2026. Across 48 answers the products displayed 185 citations that reached 136 distinct pages; thirteen were the record itself, none documented an independent check, and 90 could not be followed from what the product showed. Twenty-nine of the 48 answers reached no primary record.
The paper closes with a course scaffold: six modules, each with readings, a field lab and a deliverable, assessed against five observable competencies, with the deposited audit data as the worked example and a template for student replication.
The audit dataset, capture files, page evidence and codebook are deposited at doi:10.5281/zenodo.22839545.