Model Collapse
Training models repeatedly on their own generated data, causing diversity loss and error accumulation.
In plain language
As more web content is AI-generated, new models may learn previous models’ outputs rather than real-world data. Research shows this can erase distribution tails, homogenize outputs and overconfidence in a few patterns—model collapse. Mitigations include retaining real data, human feedback, data provenance and controlled synthetic-data ratios. For content platforms, detecting low-quality AI content and protecting original sources is infrastructure for long-term model quality.
Example
If each generation trains only on the previous generation’s paper summaries, rare viewpoints disappear and style homogenizes.
Recommended tools
Related terms
Benchmark
A standardized dataset and metric for comparing model capabilities.
Fine-tuning
Continue training a model on task data to better match a style or task.
Synthetic Data
Artificially generated data used for training or evaluation.
Agent Memory
A mechanism for an agent to store and retrieve user preferences, past tasks and key facts across sessions.