A plain-language reference on how synthetic training data, provenance, and licensing actually work — written for the teams who have to defend their data, not just use it.
Synthetic training data is data generated by a model or algorithm rather than collected from the real world. Instead of scraping video, images, or audio from existing sources, each asset is produced from a generative model to a defined specification. Because nothing is collected, a properly produced synthetic dataset contains no source footage, no scraped material, and no real-person likeness — which is what allows it to carry a clean and documented origin.
Most suppliers describe their documentation. Here is ours — the actual per-asset record from the public example manifest that ships with our open-source verifier:
{
"path": "assets/scene_0001.png",
"sha256": "c0c4002a61f865bb1c58cbcae7ef75a85a14dc6e124638ba8b9c36a8743c5162",
"bytes": 12420,
"origin": "synthetic",
"generator": {
"model": "example-noise-field",
"version": "1.0",
"license": "MIT (example)"
},
"params": { "seed": 7919, "size": "64x64" }
}
That record is not an illustration. It is the first asset in the example manifest at github.com/TheGray-Systems/tgs-verify — clone the repository, run the verifier, and it will recompute that fingerprint from the file and confirm the match. One record answers what generated the asset, under which license, with which parameters, and whether the delivered bytes are the documented bytes. Multiply it by every asset in a delivery and you have a chain of custody. This page can tell you what provenance means; that record is what it looks like when it has to hold.
No — and any provider claiming otherwise is overstating it. Synthetic data is only as defensible as its documentation and its generator. Two risks remain if they aren't addressed: the generating model itself may have been trained on copyrighted material, and a model can reproduce stylistic or structural patterns resembling protected works. Synthetic data becomes defensible when it is generated from a permissively licensed model, with documented parameters, and with prompts and outputs disciplined to avoid reproducing identifiable protected works. The safety isn't in the word “synthetic.” It's in the documentation and the discipline behind it.
No. Public visibility is not a license. Content that can be freely viewed — videos on a public platform, images on the open web, posts on social media — remains owned by its creator and governed by the platform's terms and copyright law. “Publicly available” describes access, not rights. Whether unlicensed training is defensible is being litigated right now, and the rulings are mixed — which is precisely the point: a dataset whose legality depends on the outcome of someone else’s lawsuit is a risk position, not an asset. This is the most common and most expensive confusion in training-data sourcing: teams treat viewability as permission, and the gap surfaces later, during diligence or litigation, when it is hardest to fix. The question to ask of any dataset is never “was it public?” — it is “who granted the right to train on it, and can you show me the grant?”
Model collapse is the degradation observed when models are trained recursively on undifferentiated, unfiltered synthetic output — each generation learning from the last generation's mistakes until diversity and accuracy erode. The risk is real, but it describes a practice, not a material. The collapse findings involve closed loops of uncurated self-generated data. Purpose-built synthetic data works differently: it is generated to a specification, quality-gated before delivery, documented per asset, and mixed deliberately into a training corpus rather than fed back blindly. The defense against collapse is curation — knowing exactly what each asset is, filtering what fails, and controlling the mix. That is a documentation and quality-control problem, and it is solvable. Undocumented synthetic data is where the collapse risk actually lives.
Yes — in two ways, and only one of them gets talked about. The generator model carries its own priors: leave a prompt vague and the model fills the gap with its defaults, skewing the output toward whatever it saw most in its own training. And the specification can be biased: if nobody decides the distribution — ages, skin tones, lighting, environments, accents — the dataset ends up with whatever fell out. The difference with generated-to-spec data is that the distribution is a decision instead of an accident: it is written into the specification, executed in the generation parameters, and both travel with the delivery. That doesn't make bias impossible; it makes it inspectable. A buyer can read the distribution that was ordered, check the parameters that produced it, and measure the delivered set against both. Collected data cannot offer that — its distribution is whatever the collection happened to catch. The record it demands: the distribution specification and per-asset generation parameters, so "representative" is a checkable claim instead of a marketing word.
Synthetic data wins where real data is legally encumbered, impossible to collect at scale, or dangerous to source: rights-clean visual training material, rare edge cases, controlled distributions, and content where no real-person likeness can be permitted. Real, collected data wins where authentic human variation is the point — natural speech, real accents and dialects, genuine environments and behaviors — provided it is collected with documented consent and clean rights. Most serious training pipelines need both. The deciding question is not synthetic versus real; it is whether the data's origin and rights can be proven either way. A defensible pipeline holds both kinds of data to the same evidentiary standard.
Data provenance is the documented history of where a piece of data came from and how it was produced. For training data, provenance answers: what generated this, under what license, with what parameters, and when. Provenance is the difference between claiming data is clean and proving it. Buyers increasingly treat any dataset without provenance as a black box — and a liability that surfaces during diligence.
Collected data answers provenance with a chain of ownership: where it was captured, who owned it, what rights transferred. Generated data has no such chain — nothing was filmed, scraped, or bought — so the question changes shape. Synthetic provenance answers four things instead: what model generated the asset, under what license that model was operated, with what parameters and seed, and whether the delivered bytes are the exact output — the content fingerprint. One honest boundary remains: the generator model itself was trained on something, and that history lives with the model, not the asset — which is why the record names the model and its license rather than pretending the question away; the model license is where that risk is allocated. A generated asset with those four answers documented is more traceable than most collected data will ever be: its entire causal history fits in one record. The record it demands: generator identity, version, and license; generation parameters and seed; a declaration of synthetic origin; and a SHA-256 fingerprint of the delivered bytes — per asset.
Chain of custody is an unbroken, documented record of an asset from creation to delivery, proving it is exactly what it claims to be and was never substituted or altered. Borrowed from law and forensics, it applies directly to training data: a buyer who may one day have to defend their dataset in front of an auditor, an acquirer, or a court needs evidence, not assurances. A dataset with an intact chain of custody — generator, license, parameters, content fingerprint, origin statement, per asset — is evidence that holds. A dataset with gaps in that chain is evidence that fails.
C2PA (the Coalition for Content Provenance and Authenticity) is an open technical standard for attaching cryptographically signed provenance records — often called Content Credentials — directly to a media file. A C2PA manifest travels with the asset and records how it was created and modified; because the record is cryptographically signed, any tampering with the asset or its history is detectable. For AI training data, C2PA-style signed provenance turns a supplier's origin claims into verifiable evidence: instead of trusting a spreadsheet, a buyer can validate the signature on each asset. It is the emerging benchmark for what “documented origin” should mean.
A model does not learn from raw video or images alone; it learns from the pairing of each asset with accurate, structured annotation. Thin or inaccurate metadata makes even high-quality content hard to train on, while rich, accurate metadata makes even modest content highly trainable. For training data, the annotation payload is not an accessory to the product — it is a core part of the product. This is why a documented, well-annotated dataset can outperform a visually superior but poorly documented one.
The same dataset can be delivered under different rights structures, each fitting a different buyer need. A non-exclusive training license permits use for model training while the supplier may license the same data to others. An exclusive license grants the data to one buyer only. A full IP buyout transfers complete ownership to the buyer for proprietary use. The right structure depends on the buyer's risk posture and roadmap — and because the generation cost is the same across them, the rights structure, not the pixels, is often what determines value. A fourth, narrower class — evaluation-only — covers the sample batches a buyer validates before committing to volume.
A clean synthetic dataset can be verified through its documentation: the generating model and its license, the generation parameters and seeds (which make outputs reproducible — itself evidence of synthetic origin), a content fingerprint per asset, and a clear statement of origin. Reproducibility is strong proof: if a supplier can regenerate an asset from its recorded model, parameters, and seed, the asset is demonstrably synthetic. Bit-exact replay varies by pipeline — which is why every asset also carries a content fingerprint recorded at generation: the fingerprint proves what was delivered; the parameters prove how it was made. A supplier who cannot show this within minutes is selling a black box.
Verification should never require trusting the supplier. Every TGS delivery ships with a custody manifest — a machine-readable record of each asset’s fingerprint, origin, generator, and license — and the format is public: the TGS Custody Manifest Format and its reference verifier, tgs-verify, are open source on GitHub. Point the verifier at a manifest and the delivered files; it recomputes every SHA-256 fingerprint and reports, loudly, whether the data received is exactly the data documented. The check runs offline, takes seconds, and works for anyone in the chain — buyer, counsel, auditor, or acquirer.
In July 2026, a U.S. federal court granted final approval to a $1.5 billion settlement — the largest copyright settlement in U.S. history. The court held that training AI on books was fair use. The liability came from acquiring and storing roughly 500,000 pirated works, valued at about $3,000 per work. The legal risk in training data is acquisition provenance, not training itself. The Gray Systems addresses this directly: every dataset ships with a documented chain of custody recording how each item was sourced and licensed.
A buyer evaluating a synthetic data supplier should ask: what model generated this and under what license; can you show the generation parameters and reproduce an asset; what does the per-asset metadata contain; what rights structures are available; and can you provide documentation that would survive legal diligence. A supplier who answers these clearly is selling defensible data. A supplier who deflects is selling risk.