Licensed, collected or synthetic: three ways to source training data

Every training set starts with the same decision, usually made without anyone naming it: do we buy permission to use data that exists, pay people to make data that does not, or have a machine produce it? The answer is rarely one route. It is a mix, and the mix should follow the gaps in the model rather than the habits of whoever is running procurement.

Start from the gap, not the vendor

Before talking to any of the AI training data companies in the market, write down two things: the behavior the model has to learn, and the test that will show it learned it. “Better at insurance claims” is not a gap. “Extracts the policy number, loss date and claimed amount from scanned claim letters with 98% field accuracy” is.

That sentence tells you what kind of data closes the gap. If the behavior depends on wide exposure (many writing styles, many document layouts, years of real conversations), licensed data is the obvious start. If it depends on judgment that only trained people have, someone has to be paid to produce it. If it depends on rare cases that are expensive to find in the wild, a generator may be the cheapest way to get enough of them.

Route one: license what already exists

Licensing pays for permission rather than labor. The seller already holds the material: a publisher’s archive, a forum’s posts, a stock library, a company’s support history. You negotiate what you may do with it and for how long.

What licensing is good at

Breadth. No commissioning budget recreates ten years of forum threads or a million hours of varied video. Public figures are scarce, but Bloomberg reported in January 2025 that AI companies were paying $1 to $4 a minute for unused video footage, which gives a sense of how licensed media gets priced: by volume, with the rate set by how unusual the footage is.

Where licensing goes wrong

The chain of title. A reseller can only license what its own agreement lets it sublicense, and “for internal analytics” is not “for machine learning training”. Collective licenses help with published works: CCC’s overview sheet presents its AI Systems Training License as one voluntary, non-exclusive permit, gathered from many rightsholders, for AI used outside the licensee’s own organization.

Route two: commission collection or labeling

Commissioning means a vendor’s people create or label data to your specification. That covers annotation of data you already hold, new recordings or photographs, written demonstrations, and preference rankings of your model’s answers.

What commissioning is good at

Fit. You write the guidelines, so the data teaches exactly the behavior you defined. Contracts can also give you ownership of what is produced, which licensing rarely does.

What it costs

It depends on how much skill each item needs. At the simple end, Label Your Data publishes bounding boxes from $0.02 per object. At the expert end, the people Epoch AI interviewed for its January 2026 FAQ said a single graded reinforcement learning task tends to cost between $200 and $2,000. Most vendors between those two points publish no rates at all and quote after scoping.

Kind of commissioned work	Typical unit	Published example
Image labels	Per object	Bounding boxes from $0.02 (Label Your Data)
Annotator time	Per hour	Typically $6 per annotator hour (Label Your Data)
RL tasks with graders	Per task	$200–$2,000 typical (Epoch AI interviews)

Where commissioning goes wrong

Consistency. Ten annotators read the same guideline ten ways. A paid pilot of a few hundred items, reviewed by your own team, shows you the disagreement rate before you are committed to the full volume.

Route three: generate synthetic data

Synthetic data comes from a model, a simulator or a rules engine. Tabular generators learn the shape of a real table and produce new rows; rendering platforms produce labeled images of scenes that would be dangerous or rare to photograph; a stronger language model can write instruction data for a smaller one.

What synthetic data is good at

Scale and privacy. Once a pipeline exists, the marginal item is cheap, and records that cannot leave a secure environment can be replaced with generated ones that keep the statistical shape. Pricing is usually software pricing: Rendered.ai, one of few to publish, lists its platform from $5,000 a month.

Where synthetic data goes wrong

It inherits the generator’s blind spots. If the model that wrote your training examples cannot reason about a case, neither will the model trained on them. The only reliable check is a held-out set of real data that the synthetic set never touched.

Mixing the three

Most production sets combine all three. A common shape:

1. A licensed or collected core that gives breadth and real-world variety.
2. Expert-labeled or expert-written examples for the hard cases the core gets wrong.
3. Synthetic data to fill rare classes, generated from and checked against the core.
4. A real, held-out evaluation set that none of the above ever touches.

Price each layer in the same unit, cost per accepted example, and the trade-offs become visible. The cost estimator does the arithmetic for the units that have published rates; for the rest, a scoped quote is the only honest number.

A short way to decide

• If the data exists and the license is clean, license it.
• If the behavior needs people who know the domain, commission it, and pay for a pilot first.
• If real examples are rare or sensitive, generate them, and keep real data back to test.

The route you pick matters less than whether you can show, months later, where every example came from.