19 straight answers on synthetic training data, provenance, licensing, and AI verification. Start with the question you need to settle. Inspect the records behind the answer.
Synthetic training data is data generated by a model or algorithm rather than captured directly from the real world. It can be produced to a defined specification for model training or evaluation. Generation alone does not establish that an asset is free of third-party material, likenesses, or rights restrictions. A defensible dataset documents its generator, inputs, parameters, output review, and applicable rights.
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" }
}
This is a working example from the public repository, rather than a customer delivery. 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. The example records a declared generator, license, parameters, and a checkable file fingerprint. The verifier confirms the file match; the supporting production and rights records establish the basis for the other declarations. A manifest is part of a custody record, rather than a complete log of every handling event. This page can tell you what provenance means; that record is what it looks like when it has to hold.
No. Generating an asset does not establish that it is free of third-party rights. Review the generator’s terms, relevant inputs, output restrictions, and any protected expression or likeness in the result. A permissive model license is one piece of that review; it does not independently clear the model’s training sources or every generated output. Keep the rights assessment alongside the production record.
Not automatically. Public visibility describes access, not a grant of rights. Material may be protected, licensed for particular uses, or in the public domain. Assess the actual source, applicable terms, privacy requirements, intended use, and any legal exception being relied on. Record the license or legal basis rather than treating public access as permission.
Yes. Research has shown that recursively training successive models on generated data can lose information about the original distribution. Documentation alone does not prevent this. Evaluate quality, diversity, filtering, the mix with other data, and downstream performance against an appropriate holdout set. A purpose-built synthetic dataset should be validated for the intended training task rather than assumed safe because its provenance is documented. Read the research.
Yes. Bias can enter through the generator, prompts, specification, filtering, and output selection. A written distribution target makes the intended coverage inspectable, but delivered examples still need to be measured against it. Collected datasets can also be designed and sampled to a target distribution. For either method, preserve the specification, sampling or generation records, output checks, and identified limitations.
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 an asset’s source, production, and changes. For training data, it may include the generator or collection source, relevant inputs and parameters, timestamps, and applicable rights. Provenance supports a review of origin claims; it is not a substitute for inspecting the underlying evidence or assessing the intended use.
Generated-data provenance documents the model and version, relevant inputs, generation parameters where disclosed, applicable terms, and the delivered-file fingerprint. Input assets may have their own source and rights history. A model license does not establish every fact about its training corpus. State what is known, retain the supporting records, and disclose gaps rather than describing one manifest as the asset’s entire causal history.
Chain of custody is the documented history of an asset from creation through delivery, including who handled it and any recorded changes. A content fingerprint helps check whether delivered bytes match the recorded asset. The accompanying generator, license, parameters, and origin statements provide evidence to review; a matching fingerprint does not independently prove those statements are true. Gaps, unrecorded changes, and unsupported rights claims should be disclosed.
C2PA (the Coalition for Content Provenance and Authenticity) is an open standard for cryptographically signed content provenance, often called Content Credentials. Validation checks the credential, its binding to the asset, and its signing trust chain. It does not by itself establish that every recorded assertion is true, that all ingredients are available for inspection, or that the buyer has training rights. TGS custody manifests and SHA-256 checks are a separate format; they should not be described as C2PA credentials without a conforming implementation. Read the C2PA explainer.
Metadata makes a dataset easier to inspect, filter, join, and evaluate. Useful fields depend on the task: labels, descriptions, distribution attributes, production details, and rights records serve different purposes. Not every training method requires paired annotations. Assess whether the metadata is accurate and useful for the intended pipeline; rich metadata alone does not establish better model performance.
TGS uses six delivery classes: evaluation-only, training-only, non-exclusive, exclusive, full IP buyout, and custom. The rights and restrictions depend on the written agreement for each delivery. Exclusivity may be limited by term or field of use. A buyout transfers the rights specified in the agreement to the extent those rights exist and can be transferred; production methods and pipelines are separate.
Review the generating model and its terms, relevant inputs, generation records, output checks, and the rights granted for the intended use. Recompute each asset’s fingerprint against the custody manifest to check byte integrity. Replay can support a generation claim where the pipeline permits it, but recorded parameters alone do not prove origin. Neither a matching hash nor a reproducible output establishes that the generator’s training sources or the output’s rights are clear.
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.
On July 20, 2026, a federal court granted final approval to a $1.5 billion settlement resolving claims concerning Anthropic’s acquisition of books. The case shows why acquisition records matter. It does not provide a general exemption for all AI training or settle every question about other datasets, outputs, or jurisdictions. Buyers should document both how material was acquired and the rights or legal basis for its intended use. Read class counsel’s settlement update.
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.
No. A SHA-256 match establishes that the checked file matches the recorded fingerprint. Origin and rights depend on the supporting generation or collection records, licenses, consent where relevant, and the reliability of the parties recording them. Inspect the evidence behind the manifest as well as the bytes.
Both require a record that can be inspected. Data provenance documents the asset and its history; AI behavior verification compares the expected action, the agent’s reported result, and the state observed in the destination system. Preserving a trace or screenshot establishes which evidence was retained. It does not, on its own, establish that the task succeeded. The report must state the test boundary, finding, and limitations.