The Compounding Error Cycle is the self-reinforcing loop where stale knowledge produces confident but incorrect AI output, which generates conversations that contaminate the AI's memory, which then cross-pollinates with updated knowledge to produce contradictory outputs that are progressively harder to detect and fix. Each cycle deepens the contamination across multiple knowledge layers, making the errors compound like interest on bad debt.

Feedback loops in AI systems are a recognized concern in machine learning research, from reward hacking in reinforcement learning to model collapse documented by Ilia Shumailov and colleagues at Oxford. Jason Barnard's Compounding Error Cycle (2026) identifies an equivalent feedback loop operating not at the model layer but at the knowledge layer: stale knowledge produces confident output, which generates conversations that contaminate memory, which cross-pollinates with updated knowledge to produce contradictions that are progressively harder to detect. Where model collapse describes synthetic data poisoning training data, the Compounding Error Cycle describes real conversations poisoning knowledge bases. The concept explains why AI assistants degrade non-linearly rather than gradually, and why entrepreneurs experience the frustrating rebuild-from-scratch cycle. It is the escalation mechanism within Jason Barnard's Knowledge Rot diagnostic framework, sitting alongside The Confidence Fallacy (the mask) and The Colleague Fallacy (the cause).