Short answer. Conditionally, and the condition does most of the work. Shumailov and colleagues showed in Nature in 2024 that indiscriminately training generative models on recursively generated content — each generation learning from the previous generation's output — causes model collapse: the tails of the distribution disappear first, then the output converges toward something narrow and low-variance. Synthetic data used deliberately, mixed with real data, filtered on quality and anchored to human-verified material is a normal and effective technique. Synthetic data accumulated accidentally, unlabelled, from a web that is increasingly machine-written, is the failure case — and it is the one most organisations are exposed to without ever having chosen it.
Neither synthetic nor real data is universally better. Real data captures authentic complexity — sensor noise, unexpected behaviour, naturally occurring edge cases, cultural variation — and is the reference against which synthetic fidelity is judged. It is also expensive, slow, sometimes scarce and often sensitive. Synthetic data is controllable, cheap to vary and available on demand, and it can reproduce its generator's biases while looking entirely realistic.
The question worth answering is therefore not which is better but under what conditions generated data degrades the model that learns from it. There is a published answer.
What the Nature paper actually showed
In July 2024, Shumailov, Shumaylov, Zhao, Papernot, Anderson and Gal published "AI models collapse when trained on recursively generated data" in Nature (volume 631, pages 755–759). The experimental setup is the part worth understanding precisely, because the popular summary of the result is broader than the result.
The researchers trained a model, generated data from it, trained the next generation on that output, and repeated. Across successive generations the output distribution narrowed. Low-probability events — the tails — disappeared first, because they are under-represented in any finite sample and therefore under-represented in what the next generation learns from. Over further generations the distribution converged toward something with far less variance than the original data. The authors' conclusion is that indiscriminately training generative AI on a mixture of real and generated content, as happens when data is scraped from the internet, can lead to a collapse in the models' ability to generate diverse, high-quality output.
Collapse is hard to notice because it begins in the tails. A model losing its ability to produce rare-but-valid outputs still performs well on average, on common cases, and on aggregate benchmarks. By the time a headline metric moves, a great deal of diversity has already gone — which is an argument for evaluating on rare cases specifically rather than relying on averages.
The result has also been examined critically, which is normal and useful. A note published on arXiv (preprint 2410.12954, October 2024, not peer-reviewed) argues that particular assumptions in the original experimental setup bear on how strongly the conclusion generalises. The honest practitioner position: the mechanism is real and well-motivated, its severity in any specific pipeline depends on how that pipeline curates data, and the appropriate response is curation rather than either panic or dismissal.
When is synthetic data the right choice?
Synthetic data is not a degraded substitute for real data in every application. There are cases where it is straightforwardly better, and reading the collapse result as a blanket prohibition gives them up for no benefit.
| Use case | Why synthetic works | The condition |
|---|---|---|
| Rare events — accidents, faults, hazards | Real examples are scarce, dangerous or unethical to collect | Validate against the real examples that do exist; never let synthetic define the category alone |
| Privacy-sensitive domains | Removes personal records from the training set | Verify the synthesis does not leak identifiable attributes from its source |
| Class balancing | Cheap augmentation of under-represented classes | Cap the synthetic share and check whether the minority class behaviour actually improves |
| Simulation for perception and robotics | Perfect ground truth, unlimited variation, controllable conditions | Measure the sim-to-real gap explicitly and close it with real validation data |
| Distillation from a stronger model | Deliberate transfer from a known, higher-quality teacher | Not recursive self-training — the teacher is better than the student and the lineage is known |
| Formats and structure | Templates and structured variation are cheap to generate correctly | Content still needs human grounding; structure is not substance |
The pattern across the right-hand column: synthetic data is safe when its lineage is known and it is anchored to something real. It is dangerous when it is unlabelled, accumulated, and validated only against other synthetic data.
One assumption worth retiring while you are here. Synthetic does not automatically mean private. NIST has emphasised that many synthetic-data techniques provide no formal privacy guarantee, so when privacy is the objective, the generation method and its residual risk have to be evaluated rather than assumed away because the records are artificial.
The exposure nobody chose
Most organisations reading about model collapse conclude it does not apply to them because they do not train foundation models. The exposure is more indirect, and more common, than that.
- Scraped corpora. Any dataset assembled from the open web after roughly 2023 contains machine-generated text in unknown proportion, unlabelled. Fine-tuning on it is recursive training whether or not that was the intention.
- Fine-tuning on your own outputs. Teams that publish AI-assisted content and later fine-tune on their own published corpus have built a small recursive loop with no external anchor.
- Evaluation sets contaminated by generation. Benchmarks assembled from web text may contain generated items, which flatters models producing similar text.
- Annotation anchored to model suggestions. Model-assisted labelling accepted without unassisted control batches encodes the pre-labelling model's distribution into the dataset — the same mechanism reached from a different direction.
- Retrieval corpora. A retrieval-augmented system whose index is full of machine-written pages retrieves machine-written answers, with no training involved at all.
The common factor is missing provenance. None of these is dangerous if you know which items are human-authored and which are not, because then they can be weighted, filtered or excluded. All of them are dangerous when that information was never recorded — which is a practical argument for provenance discipline entirely separate from the legal one.
A related trap sits in evaluation: if the same family of models both generates the data and judges it, the system rewards its own assumptions, and the circularity is invisible from inside the scores.
Using synthetic data without degrading the model
The objective is deliberate composition with known lineage, rather than either avoiding synthetic data or accumulating it unnoticed.
- Label every item with its origin. Human-authored, model-generated, model-assisted, or unknown. "Unknown" is a legitimate and important category — track it, rather than silently treating it as human.
- Anchor the mix to verified human data. Keep a substantial, curated core of human-authored, human-verified material that does not shrink as synthetic volume grows. This is what stops distributional drift, and its value rises as the open web becomes less reliable as a source.
- Set and enforce a synthetic share. Decide the proportion deliberately, per training run and per domain, and record it. A mix nobody chose is a mix nobody can debug when behaviour changes.
- Filter on quality, not just validity. Generated items that are well-formed but generic are the ones that drive distributional narrowing. Filtering for diversity and informativeness matters more than filtering for correctness alone.
- Evaluate on the tails. Build evaluation sets specifically from rare cases, minority classes and edge conditions. Aggregate metrics stay healthy through the early stages of collapse; rare-case performance does not.
- Keep an untouched real-data holdout. A human-authored evaluation set, collected once, never used for training, never regenerated and never augmented. It is the only stable reference point across model generations.
- Version the data, not only the model. Record which mix produced which checkpoint. When behaviour degrades, the data composition is usually the explanation, and it is unrecoverable if it was not recorded.
Synthetic share = Machine-generated items ÷ Total items in the training mix
The right ratio is determined experimentally for the target task, not by a universal rule — and it should be reported alongside segment-level performance on the segments the synthetic data was added to improve. If those segments did not move, the augmentation did nothing except change the distribution.
What this implies for anyone publishing at volume
There is a second-order implication, and it is uncomfortable for the content industry specifically. If the open web fills with machine-generated material, the corpus future models learn from degrades, and the organisations producing that material are contributing to the degradation. Publishing fluent, generic, unverified content at scale is not only a weak marketing asset; it is a small contribution to a collective problem.
The constructive reading is that this raises the value of two things producers control. First, human-verified substance: original data, first-hand experience, genuine expertise, real in-market linguistic work — material that remains valuable precisely because it is not a resample of what already exists. Second, provenance marking, which lets downstream curators distinguish what you generated from what you wrote, at almost no cost at export time.
How Lifewood approaches this
Lifewood sits on both sides of this problem: it collects and annotates human-authored training data, and it produces AI-generated content under human direction. The rule applied in both directions is the same — origin is recorded rather than assumed, and human verification is the anchor, held to a 95%+ accuracy threshold under dual-layer human-in-the-loop review. The reason that anchor is producible at all is the network behind it: 50+ languages, 40+ delivery centres across 30+ countries and 56,788 contributors, which is what it takes to generate genuinely new human material rather than recycled web text.
See global AI data, AI data validation, what to buy: RLHF, SFT or distillation for the deliberate-teacher case, and AIGC governance, disclosure and provenance.
Sources and further reading
- Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R. and Gal, Y., "AI models collapse when trained on recursively generated data", Nature 631, 755–759, July 2024.
- "A Note on Shumailov et al. (2024)", arXiv preprint 2410.12954, October 2024 — critical commentary, not peer-reviewed.
- NIST, Differentially Private Synthetic Data — on why synthetic records are not privacy-preserving by construction.