Model collapse is a degenerative failure mode in which a generative model trained recursively on data produced by earlier generations of models progressively forgets the tails of the true data distribution, converging toward a narrow, high-likelihood subset. It is distinct from training on synthetic data in general, since well-curated synthetic data can improve models; collapse is the specific recursive case in which model outputs feed future training without sufficient grounding in real data.
Definition
Two variants are usually distinguished. In early model collapse, tails and rare events disappear while overall quality on common inputs still looks fine. In late model collapse, modes merge, the distribution contracts toward a near-delta, and outputs become repetitive, low-entropy, and factually unmoored.
Mechanism
The process unfolds across model generations in four stages. First, a generation-N model is trained on a mix of real and synthetic data. Second, its outputs systematically under-represent rare events such as heavy-tail phenomena, minority dialects, and unusual facts, because low-likelihood sampling is rare. Third, generation-N+1 trains on a corpus containing more model-generated content, inheriting and amplifying the truncation. Fourth, over successive generations, tails collapse, modes merge, and the model's distribution drifts away from the real-world data-generating process.
The effect has two distinct sources. The first, statistical approximation error, arises because finite samples under-represent tails, and this compounds over generations. The second, functional approximation error, arises because model architectures cannot perfectly represent the true distribution, so approximation error likewise compounds.
Origin papers
Shumailov et al., "The Curse of Recursion: Training on Generated Data Makes Models Forget," 2023 (later published in Nature as "AI models collapse when trained on recursively generated data," 2024), is the canonical reference and coined the term "model collapse." Alemohammad et al., "Self-Consuming Generative Models Go MAD," 2023, gave a parallel demonstration in image generation, terming the phenomenon "Model Autophagy Disorder." Gerstgrasser et al., "Is Model Collapse Inevitable?," 2024, argued that accumulating real data alongside synthetic data, rather than replacing it, prevents collapse.
Mitigations
Several mitigation strategies have been proposed. Accumulation rather than replacement keeps adding real data instead of rolling it forward; Gerstgrasser et al. show accumulation prevents asymptotic collapse. Data mixing bounds the fraction of synthetic data in any training run. Provenance tagging, including watermarking and C2PA labels, allows training pipelines to identify and down-weight synthetic content. Diverse teachers, via multi-teacher distillation, reduce mode collapse relative to single-teacher recursion. Anti-contamination filters, including perplexity filters, classifier detectors, and domain heuristics, attempt to strip model-generated content from web crawls.
Policy and safety relevance
As the web fills with AI-generated content, future model training may be increasingly affected even without deliberate recursive training, a data-pollution concern that is a central argument for provenance standards and watermarking. Disclosure regimes such as the EU AI Act GPAI rules and the EU General-Purpose AI Code of Practice (Final Version, 2025) implicitly assume that training data composition is knowable; model collapse makes composition matter for public-interest reasons, not just copyright. There is also a monoculture risk: if one teacher model seeds most synthetic data industry-wide, its biases and blindspots propagate to every downstream model. Some regulators, including France's CNIL and the UK's CDEI, have flagged output-diversity loss as a consumer-harm axis.
Current state
Empirical evidence of collapse in real frontier pipelines is limited, largely because labs carefully curate synthetic mixes. The theoretical argument is uncontested; the practical question is how much real data is enough to avoid collapse. Web-scale contamination is the larger concern: by some estimates, a meaningful fraction of recent web text is AI-generated, which makes the real-data accumulation path itself leaky.
Relationships
- contradicts (partially): Synthetic Data / Model-Generated Training Data — model collapse is the failure mode when synthetic data is used recursively without mitigation
- depends-on: Data Provenance, C2PA, and Watermarking — provenance is the main structural mitigation
- related: Distillation — single-teacher recursion is a collapse-adjacent pattern
- related: AI Software Progress — if collapse is real at scale, measured efficiency gains may decelerate