Glossary
Updated
Describes both intentional (reinforcement learning) and unintentional (recommendation reinforcement) loops. Positive loops can optimize performance over time, but negative loops risk bias amplification (e.g., a recommender showing popular content makes it more popular). Governance strategies include loop-detection metrics, intervention policies (diversity quotas), and simulated-loop testing before live deployment.
Example
A news platform’s recommender shows trending articles. Because users click more on those, the system further amplifies them, narrowing content diversity. The team introduces “serendipity” constraints (injecting less-clicked topics at a fixed rate) to break the unbounded feedback loop and maintain content variety.