Model collapse is a failure process in machine learning in which models trained repeatedly on data generated by earlier models lose information about the original data distribution. Successive generations may underrepresent rare events, produce increasingly homogeneous outputs, or develop statistical distortions that reduce their usefulness. The concept is especially relevant to generative artificial intelligence, including language and image generation, but it can also arise in simpler statistical models. It does not mean that every use of generated training data causes deterioration: outcomes depend on how data are generated, selected, retained, and combined with independently obtained observations.
Definition and scope
A generative model learns an approximation to a probability distribution from training data. Its outputs can then become training examples for another model, either deliberately or because generated material enters a larger data collection. When this process repeats, inaccuracies in one generation can influence the data available to subsequent generations.
Model collapse concerns the resulting loss of fidelity to the original distribution. “Collapse” need not mean an abrupt failure or completely meaningless output. A model may remain fluent and perform adequately on common examples while losing coverage of unusual cases. More severe deterioration may eventually affect ordinary examples as well.
The term generally describes a process across successive training rounds or model generations, rather than an isolated low-quality response. The models need not have identical architectures, and a model need not literally train on its own outputs. Training on outputs from other models can create a similar feedback loop.
Relevant systems include large language models, diffusion models, and other learned generators. Their failure patterns differ, so model collapse is better understood as a family of related phenomena than as a single universal trajectory.
How recursive training causes information loss
Finite sampling and rare events
Generated datasets are finite samples, not complete representations of a model’s learned distribution. Rare events can therefore disappear even when the generator represents them reasonably well.
If an event has probability , its probability of appearing zero times in independent samples is
For sufficiently small , omission remains likely even in a substantial dataset. If the next model learns only from that dataset, it may assign the event a still smaller probability or fail to represent it altogether. Repeated sampling and retraining can compound this loss.
This mechanism helps explain why low-probability regions—the tails of a distribution—can deteriorate before common patterns visibly change. Once information disappears from both the model and its training material, additional generations of sampling do not reliably reconstruct it.
Approximation and training error
Sampling is only one source of distortion. A model may not represent the target distribution exactly because of limited capacity, its inductive bias, imperfect optimization, or the chosen loss function. The learned distribution therefore differs from the source distribution before generated examples are sampled.
In recursive training, these discrepancies influence the next generation’s input. Some errors may diminish, while others persist or grow. There is no general rule that every error increases monotonically, but repeated training on imperfect approximations can progressively move the model away from the original data.
Generation and selection choices
The distribution of generated examples also depends on how outputs are produced. Decoding settings, prompts, rejection rules, and quality filters can favor some outputs over others.
For example, beam search can emphasize high-probability sequences rather than provide representative samples from a language model. Filtering for conventional, polished answers may improve individual examples while narrowing the range of styles or viewpoints in the resulting dataset.
Consequently, collapse is not determined only by the proportion of generated data. The diversity, coverage, and selection process of that data also matter.
A simplified mathematical model
Let denote an original data distribution, and let denote the distribution represented by a model at generation . A purely recursive procedure can be written schematically as
where contains independently generated examples and is a training procedure.
In an idealized setting with unlimited representative data, an exactly expressive model, and exact training, a distribution could reproduce itself. Practical recursive training departs from those assumptions.
A simple illustration uses a model of a binary outcome. Suppose its current probability of one outcome is . Generate observations and estimate the next probability from their observed frequency:
The conditional expected value remains unchanged:
but the conditional variance is
Thus, an estimator can be unbiased at each step while individual training lineages drift. The states and are absorbing: after either outcome disappears, this procedure cannot generate it again. With a fixed finite sample size, the process eventually reaches one of those states with probability one.
This example illustrates information loss from repeated finite sampling. It is not a complete model of neural-network training, where generalization, smoothing, external data, and training constraints can alter the dynamics.
Training regimes and their differences
Different experimental designs can produce substantially different results.
- Complete replacement: Each new model trains only on outputs from an earlier model. This directly exposes recursive information loss.
- Mixed training: Each generation receives both generated examples and independently sourced data. The continuing external input can preserve information that would otherwise disappear.
- Data accumulation: Earlier datasets are retained while generated examples are added. Original evidence remains available, unlike in complete replacement, although its relative training weight may decline.
- Verified synthetic training: Generated examples are checked using external evidence, executable tests, or other validation mechanisms before inclusion.
A simplified mixture is
where is the next training distribution and is the generated-data share. This expression omits finite-sample effects, selection biases, and changes in the external source. It nevertheless shows why replacing all original data differs from repeatedly supplying an external reference distribution.
There is no universal synthetic-data percentage at which collapse begins. The result depends on data quality, task, model, sample size, filtering, and training procedure.
Research history
Model collapse became a prominent research topic in 2023 amid concern that generated material could enter future training corpora. A widely discussed study initially circulated under the title The Curse of Recursion: Training on Generated Data Makes Models Forget. Its peer-reviewed version, AI models collapse when trained on recursively generated data, appeared in Nature in 2024.
That work combined theoretical analysis with experiments and distinguished early loss of distributional tails from more extensive deterioration. Other research investigated related recursive-generation effects, including experiments with image models.
Subsequent studies emphasized the importance of experimental assumptions. In particular, retaining original data while accumulating generated examples can behave differently from repeatedly replacing the training set. These findings limit the claim that generated data inevitably cause collapse, without removing the demonstrated risks of particular recursive training procedures.
Distinction from related problems
Mode collapse. In generative adversarial networks, mode collapse refers to a generator producing a limited subset of the target distribution’s modes. It can occur during a single training run. Model collapse concerns deterioration associated with generated-data feedback, although reduced diversity can be a symptom of both.
Overfitting. Overfitting occurs when a model fits training-specific patterns at the expense of performance on new examples. Model collapse concerns the degradation of the distribution supplied through successive training rounds. Either can occur without the other.
Catastrophic forgetting. Catastrophic forgetting is the loss of previously learned capabilities during further training. Collapse instead emphasizes information lost or distorted through generated training data, although both processes can affect the same system.
Hallucination. An AI hallucination is an unsupported or incorrect generated claim. Collapse may propagate such errors, but an individual hallucination does not establish that collapse has occurred.
Knowledge distillation. Knowledge distillation deliberately transfers behavior from a teacher model to a student. It is not inherently collapse: its success depends on the target task and training design, rather than on reproducing every feature of an original distribution indefinitely.
Evaluation and practical significance
Detecting collapse requires more than inspecting whether outputs appear fluent or plausible. Relevant measurements include:
- performance on independent, real-world test sets;
- coverage of rare events, subgroups, and unusual combinations;
- changes in output diversity and distributional fidelity;
- task-specific accuracy across successive generations;
- sensitivity to the amount and provenance of generated training material.
For language models, perplexity can provide useful evidence, but it is not a complete measure of factual reliability or diversity. Evaluation against a recursively generated test distribution can also conceal deterioration relative to the original target. Benchmark contamination is a separate concern that may make measured performance misleading.
Research on mitigation examines retaining representative source data, adding fresh observations, controlling sampling and filtering, and checking generated examples against external constraints. In programming tasks, executable tests can reject some incorrect synthetic examples, although passing tests does not establish complete correctness or adequate diversity.
Data lineage is important because recursive dependencies are difficult to study when the origins of training examples are unknown. Reliable provenance can distinguish independently collected observations from material generated by earlier models.
The broader concern is that generated content on the World Wide Web could enter later training datasets and create an indirect feedback loop. However, controlled recursive-training experiments do not by themselves establish that a particular deployed model is collapsing. Real training pipelines differ in data retention, filtering, verification, and access to new observations; these differences are central to assessing the phenomenon.