Synthetic data is not one training policy

A proposal stating that a corpus is forty per cent synthetic leaves important review questions unanswered. The generator, filtering method and hold-out result matter, but so does the sequence of training runs. The percentage describes one mixture, rather than the policy that produces successive generations.

Generated examples can supplement a versioned source corpus, replace it or accumulate in storage while a fixed training budget samples progressively fewer originals. These arrangements can share the same “synthetic augmentation” label. They expose the model to different data and should be evaluated separately.

Approve the retention and sampling policy alongside the generated material. The dataset owner needs a record of which examples remain eligible, which enter each training run and what verification they passed. That record makes the recursive-training risk inspectable.

Replacement compounds what the generator missed

A generative model is an estimate of the distribution it learned from. Samples from that model carry its approximation error. When those samples become the whole corpus for the next model, another round of estimation starts from that imperfect projection. Repeating the cycle can amplify small errors and remove low-probability parts of the original distribution.

Shumailov and colleagues call this model collapse. Their experiments across language models, variational autoencoders and Gaussian mixture models found that indiscriminate recursive use of model-generated content caused models to forget parts of the original distribution, with the tails disappearing first. Their central warning is also a data-management condition: access to the original data distribution matters through the sequence, and provenance is needed to distinguish generated material from other sources. Nature, 2024

That result is sometimes compressed into a rule that synthetic data poisons a model. The compression removes the experimental policy. Gerstgrasser and colleagues compared replacing each corpus with the latest generated data against accumulating successive generated batches alongside the original data. In their controlled experiments, replacement tended towards collapse while accumulation kept error bounded across the model families and settings tested. Their linear model analysis reaches the same distinction, although it does not turn the experimental finding into evidence about every production system. Gerstgrasser and colleagues, 2024

What the inventory says	What the training sequence does
Synthetic data replaces the source corpus	Each generation learns only from the previous model’s samples
Synthetic data accumulates beside the source corpus	Original and generated examples remain available to later runs
All generations are retained under a fixed training budget	Availability grows, but the sampler decides what the model actually sees
Synthetic share is reported as one percentage	Lineage, generation and sampling weight are hidden inside the total

The distinction explains why two studies can report different directions without one of them making synthetic data universally safe or unsafe. “Contains generated examples” describes material. It does not describe the process that made one generation evidence for the next.

Retention without sampling is only an archive

Keeping the original corpus is not enough if the training process stops drawing from it. Kazdan and colleagues tested three workflows across Gaussian estimation, kernel density estimation and language-model fine-tuning. Replacing the source data produced collapse in the settings studied. Training on all accumulated source and generated data remained stable. When data still accumulated but each generation trained on a fixed-size subset, performance degraded slowly rather than failing explosively. Kazdan and colleagues, 2024

A fixed training budget makes the sampling policy particularly important. Retaining the originals establishes their availability, not their influence on later runs. As generated batches grow, the probability of selecting an older example can fall unless sampling rules protect its intended share. Record and evaluate those rules under the actual compute budget.

Verification adds another decision. Feng and colleagues studied selection of generated examples with human or machine verifiers. In two practical tasks and several theoretical settings, they found that even imperfect verification could prevent the collapse observed when generated data was accepted without that gate. The experiments do not establish a universal verifier threshold. They do show why “filtered” needs a measurable acceptance rule and a checked relationship to downstream performance. Feng and colleagues, 2024

Approve a corpus policy, not a percentage

Require a manifest for each training run. Identify the eligible source version and generated batches, the sample drawn from each group and the verification conditions applied. Derive the aggregate synthetic percentage from those records. Use the manifest to compare policies and investigate a later result.

The holdout has a different job from the retained source corpus. Retained source examples influence training and can interrupt a recursive feedback loop. A holdout stays outside it and measures what happened. Combining the two makes the reported result optimistic and removes the independent evidence needed to test the policy.

The lineage fields matter even when the synthetic examples are good. Without them, a later team cannot reconstruct which model trained on which ancestor, cannot compare policies and cannot remove a defective generated batch without discarding unrelated material. This extends the earlier case that training-data provenance belongs in a register. That article concerns rights and acquisition. Here provenance records a statistical dependency across model generations.

illustrative fixture

The training policy defines the evidence boundary

Swipe or scroll for the full diagram →

Retention, sampling and verification must be visible across training generations.

Observed

The cited controlled studies report different behaviour under replacement, accumulation, fixed-budget sampling and verification.

Inference

Record those choices in the enterprise training manifest and evaluate the intended policy separately.

Evidence boundary

These experiments do not establish a universal collapse rate, safe synthetic percentage or protection for every task and architecture.

What this does not tell you

These studies use controlled recursive-training settings. They do not establish that an enterprise fine-tune containing generated examples will collapse, or that accumulation will protect every task, model architecture and sampling scheme. One of the accumulation experiments itself reports deterioration in a fixed-size sampling workflow. The operational proposal is to make that workflow visible and testable, not to infer a safe recipe from one paper.

Nor is source data automatically correct, representative or lawful to retain. Human-created records carry selection effects, label errors and historical policy choices. Contractual, copyright and data-protection duties can require material to be removed. When that happens, the manifest should record the change and the evaluation should test its consequence. “Original” identifies a lineage position. It is not a quality certificate.

Generated data can support coverage, simulation and tasks with checkable answers. Its suitability depends on the intended use and evaluation. Before approving the next training run, the data owner should require a defined retention, sampling and verification policy, with evidence on consequential and rare cases. A percentage without that policy leaves the training process insufficiently specified.