由 admin 
A synthetic dataset is not automatically cheap, anonymous or useful merely because its records were generated rather than observed. Procurement teams need a valuation method that records what the data helps a buyer test, which risks remain, and which obligations a seller actually accepts. This research note proposes an illustrative, auditable price model rather than a market-rate claim.
Define the unit being licensed
Start with the buyer’s permitted task: testing an interface, training a classifier, load-testing a pipeline or benchmarking rare events. These are different products even if each contains one million rows. Define schema version, generation method, number of usable scenarios, format, allowed uses, refresh period and access mechanism. A buyer paying for a repeatable benchmark needs a fixed seed and stable holdout set; a buyer testing resilience needs controlled variability. Contract language should say whether the delivered item is synthetic outputs, generator access or an ongoing service. Price cannot be compared fairly across these three units. Record the creator’s rights to the input material and the buyer’s right to redistribute derived outputs separately from the nominal row count.
Separate fidelity from coverage
A generated dataset can resemble an aggregate distribution while failing the edge cases for which the buyer paid. Construct a coverage matrix with rows for rare classes, critical relationships, missing-value patterns and temporal behavior. Require named acceptance tests instead of a single blended score. For an illustrative binary classification project, the buyer might score schema validity at 30%, critical-class recall at 40%, distributional fit at 20% and reproducibility at 10%. Those weights are design assumptions, not universal standards. A dataset that misses a safety-critical class should fail a mandatory gate even if its weighted total appears high. Document the holdout reference and its provenance to avoid circular validation.
Price the production and verification work
An example unit-economics model is total quote = generation labor + licensed compute + independent quality review + controlled delivery + support reserve + contribution margin. Assume, purely for illustration, $1,400 for generation and curation, $450 for compute, $900 for validation, $300 for delivery and $350 for support. Direct and reserved cost is $3,400. At a targeted 25% contribution margin on revenue, the modeled quote is $3,400 / 0.75 = $4,533, before taxes and negotiated rights. Do not add a second ‘risk premium’ if validation and support reserves already include the same expense. Capture hours and test logs so the margin can be recalculated when the generation method changes.
Treat privacy as a cost and acceptance gate
Synthetic is not synonymous with anonymous. Memorized examples, rare combinations and linkage with external datasets can create re-identification risk. The European Data Protection Board’s January 2025 pseudonymisation guidance makes clear that reducing linkability requires technical and organizational controls; pseudonymised personal data remains within data-protection obligations. Review training-data authorization, membership-inference tests where relevant, small-group disclosure, access controls and deletion support. Set a separate fail-closed privacy acceptance criterion instead of ‘discounting’ unlawful disclosure into the asking price. A lower price cannot cure missing lawful grounds or absent rights. Retain evidence about tests performed and blind spots rather than asserting that a generated dataset carries zero residual risk.
Include drift, refresh and warranty boundaries
A one-off dataset ages differently from a monthly regenerated feed. Define drift triggers such as schema changes, population shifts, operational rule changes or documented task-error degradation. Offer a baseline version and explicitly priced refresh cycles; do not promise unlimited updates under a static license. A sample annual offer might include four scheduled refreshes, one disputed-result investigation and a maximum response time for failed schema validation. Specify what counts as seller fault versus buyer-environment change. Track a changelog and hashes of delivered versions. The warranty should cover agreed tests and services, not a guarantee that every deployed model will perform to a promised accuracy. A useful contract maps each claim to a reproducible test case.
Run a decision table before signing
Compare vendors on accepted task coverage, measured validation labor, proof of input rights, residual privacy risk, delivery reliability and the buyer’s total integration costs. For three candidate datasets, request the same two-week pilot and split results into pass, conditional pass and reject. Use a reserved holdout and prohibit vendors from quietly training against it. Record support tickets and whether each defect can be reproduced. NIST’s July 2024 Generative AI Profile emphasizes risks around privacy, information integrity and upstream dependencies; use it as a risk vocabulary, not a certification badge. The decision rule should let a more expensive but repeatably validated dataset beat a cheap feed whose apparent savings disappear during rework.
The most defensible synthetic-data price is a traceable service quote attached to a scoped license, documented verification and an honest statement of residual uncertainty. Investors and buyers can change model inputs without pretending that invented illustrative figures are observed marketplace prices.
Sources
- EDPB, Guidelines 01/2025 on Pseudonymisation (17 January 2025, consultation draft)
- NIST, Generative AI Profile AI 600-1 (July 2024)
- NIST, AI Risk Management Framework 1.0 (26 January 2023)