# Synthetic Data Feedback Loops and AI Model Collapse

## Public-safe source report summary

This uploaded source report is preserved as durable project evidence for Cognivirus.com. It contributes concepts to the v1.15.0 danger-model expansion: Synthetic feedback loops, model collapse, variance loss.

## Evidence handling

This is treated as a **source dossier**, not as independently verified empirical consensus. Public pages may use it after applying the site evidence ladder, metaphor boundaries, and non-operational safety policy. It must not be used to claim that AI systems are conscious, literal biological viruses, or inevitably catastrophic.

## Concepts extracted for the site

- The unsafe unit may be a transition graph rather than one model artifact.
- Local component approval does not prove runtime-composition safety.
- Evidence should name the exact carrier, route, memory state, evaluator, tool profile, and promotion rule involved.
- Observable outcomes need replayable traces rather than trust language.
- Retirement, rollback, and behavioral-extinction reviews must include data, memory, synthetic examples, descendants, aliases, and human workflows.

## Source orientation

Synthetic Data Feedback Loops and AI Model Collapse Executive Summary: Modern generative AI systems increasingly train on data that was itself generated by other AI. Experts warn that this creates a synthetic feedback loop : models repeatedly learn from their own outputs. Over time this degrades performance in a phenomenon known as model collapse – the model “forgets” rare or nuanced information and produces generic, error-prone results. The effect has been likened to a viral pattern or “cognivirus,” where distortions persist and spread through the AI ecosystem. Studies show that as more web content becomes AI-generated (e.g. nearly half of n

## Site interpretation

The report is used to deepen public and technical explanations of distributed behavioral persistence, synthetic-feedback risk, action-layer controls, observability, lineage, diversity, promotion pressure, and retirement failure. It does not authorize exploit instructions, self-replication recipes, credential workflows, or backdoor construction guidance.