--- language: en title: "Model collapse" description: "Model collapse is the degradation studied in recursive training with generated data. It is not equivalent to an isolated error or to every use of synthetic data." canonical: "https://tropk.ai/en/resources/glossary/model-collapse" last_modified: "2026-10-12T16:09:31.955Z" --- # Model collapse Reviewed on 2026-10-07. ## What it means Model collapse is a process of degradation across generations of generative models when their outputs are used to train subsequent generations. Research describes a progressive loss of information about the original distribution, including less frequent cases. The problem involves how data is selected and reused, not merely the presence of AI-generated text. ## Why it matters for marketing For marketing, the concept helps explain why content volume cannot replace verifiable information. A collection formed from rewrites of rewrites can impoverish the diversity of references. This does not make it possible to diagnose a commercial assistant’s training from the response received; the team knows its own collection and sources, but usually cannot observe the provider’s internal data. ## A practical example Imagine an organization preparing a generative model to answer customer questions. With each update, it replaces documented responses with examples produced by the previous version. The team notices that rare delivery situations are disappearing from the sample and revises the data composition. It preserves documented responses for these situations and tests performance, without calling an incorrect response proven model collapse. ## How to apply or verify it In projects under your control, record the origin, proportion and selection criteria for generated and real data. Maintain independent evaluations with common and less frequent cases. Compare generations under stated conditions and investigate losses in coverage. When evaluating an external provider, request documented information about quality and provenance, without inferring a training mechanism from isolated symptoms. ## Important distinctions Model collapse differs from an isolated [hallucination](/en/resources/glossary/ai-hallucination), outdated content and deliberate contamination of a dataset. [Synthetic data](/en/resources/glossary/synthetic-data) are not necessarily useless: curation, mixing and training protocols matter. Studies define specific conditions for recursive reuse and accumulated errors. The concept describes a technical risk in model development, not an inevitable prediction about every tool or a quarterly visibility indicator. ## Sources and references - [AI models collapse when trained on recursively generated data]() — Nature ## Continue a leitura - [AI slop]() - [Circular sourcing]() - [Synthetic data]() ## About the author [Vinicius Pacheco]() Responsible for Tropk’s technical content, focusing on crawling, indexing, structured data, and agent access to content. --- [All pages in Markdown]()