Model Collapse
mechanismAI & Compute
AI that trains on AI-made content slowly forgets the rare and unusual, collapsing toward bland sameness.
Who they are
Model Collapse, the breakdown that happens when AI systems are trained on data generated by other AI.
What they do
The engine treats it as an entropy-minimizing force that flattens machine output toward uniformity.
How it works
When models train on synthetic data they lose the rare 'tail' cases through three compounding errors (statistical, expressivity, and approximation), eventually converging to a narrow, repetitive output; the engine pairs this machine-side homogenization with the human-side flattening caused by people offloading thinking to AI — two sides of one standardizing engine.
Why it matters
It matters because it points to AI and its users converging on the same narrowed, homogenized thinking across the whole network.
The engine's record — word for word
AI training on synthetic data loses tail distributions. Three mechanisms: Statistical Approximation Error, Functional Expressivity Error, Functional Approximation Error. Converges to delta function. [Report #141] HUMAN<->MACHINE dual homogenization (Report #141): recursive training of LLMs on synthetic data (this node) homogenizes machine output the way AI cognitive-offloading (cognitive_offloading_ai) homogenizes human thought — one entropy-minimization engine standardizing cognition across the network. Pairs with inverse_scale_law (scaling paradoxically degrades performance on distractor/strong-prior tasks) and algo_homogeneity.
Follow the trail
Walk this on the live map →