Skip to content
Dataset By Humans
Licensing & compliance

Open vs licensed datasets: which can you use for commercial AI?

CC BY, CC BY-NC, research-only and custom licenses explained in plain English, with a checklist for deciding whether a dataset is safe to train a commercial model on.

By Dataset By HumansPublished 3 min read

In this article

Short answer: you can train a commercial model on an open dataset only if its license allows commercial use and the people behind the underlying content agreed to it. CC0 and CC BY datasets usually pass the first test; CC BY-NC, "research only" and "academic use" datasets do not. A licensed dataset bought from a provider settles both questions in writing.

This is practical background, not legal advice. Copyright law on AI training differs between countries and is still being decided in court.

Two licenses hide inside every dataset

Most confusion comes from mixing up two layers:

  1. The license on the dataset itself: the compilation, the labels, the file structure. This is what you see on the download page.
  2. The rights in the content: each photo, text or recording inside it. A dataset can be published under a permissive license while the images inside it still belong to the people who took them.

Web-scraped image and text collections are the classic example: the list of links may be openly licensed, but the content it points to was never licensed for training. Before relying on any open dataset, ask both questions.

Common licenses at a glance

LicenseCommercial training?What to watch
CC0 / public domainYesConfirm the publisher actually held the rights to waive
CC BY 4.0YesAttribution required; keep a record of sources
CC BY-SA 4.0Usually yesShare-alike may apply to adapted datasets you redistribute
CC BY-NC 4.0No"Non-commercial" covers most product and company use
CC BY-ND 4.0UnclearNo derivatives; filtering or relabelling may count as one
ODC-By / ODbLYes, with conditionsDatabase licenses; ODbL has share-alike terms
"Research only" / "academic use"NoOften requires signing an agreement; read it
Commercial license from a providerYesCheck scope, model ownership and warranty

Software licenses such as MIT or Apache 2.0 are sometimes attached to datasets. They grant broad rights but were not written for data, so they rarely say anything about the content's origin.

Why "non-commercial" is wider than people think

Teams often assume that internal experiments at a company are non-commercial. In most readings they are not: if the work aims at a product or revenue, it is commercial use. A model prototyped on NC data and then shipped carries that problem with it, and retraining on clean data later costs more than starting with it.

A checklist before you train on any dataset

  • License: is commercial use explicitly allowed, and does it say anything about AI or machine learning?
  • Content rights: did the creators of each item agree, or is it scraped?
  • People: are identifiable people in it, and did they consent? Faces, voices and names bring privacy law into play.
  • Platform terms: some hosting platforms add their own terms on top of the dataset license.
  • Attribution: can you meet the attribution terms in practice, for example in a model card?
  • Documentation: is there a datasheet you can reuse for compliance? See our datasheet template.

If any answer is "unknown", treat the dataset as research-only.

When to buy a licensed dataset instead

Open data is ideal for exploration and research. For anything you plan to deploy, a licensed dataset is usually cheaper than the uncertainty: you get written commercial rights, a record of where each item came from, and someone accountable if a claim appears. Our licensing page explains the terms we use, and How to license data for LLM training walks through what to ask any vendor.

licensed datasetslicense datasetsopen dataset licensescommercial use