Dataset licenses: open, commercial and AI training terms

Every dataset that changes hands comes with a license, written down or not. The license says who may use the data, for what, and on what conditions. In AI it is often missing or wrong: an audit of popular training datasets by the Data Provenance Initiative found licenses missing more than 70% of the time on widely used hosting sites, with error rates above 50% (Longpre et al., arXiv, October 2023).

If you plan to license your company's data, the choice is mostly between publishing it under an open license, which lets anyone use it, and keeping it under a commercial license with a specific buyer. This guide covers both.

Open dataset licenses

Non-commercial licenses such as CC BY-NC add a ban on commercial use. They are common on research datasets and a frequent source of trouble, because whether a particular AI project counts as commercial is often unclear.

Open licenses also have a limit that matters for AI. Creative Commons notes that its licenses can't prohibit a use that copyright exceptions already permit, and that training AI on copyrighted works may be protected by those exceptions in many cases (Creative Commons, August 18, 2023). An open license grants permission; it doesn't add restrictions the law wouldn't otherwise impose.

Why open licenses rarely fit data you want paid for

An open license is permanent and available to everyone. Once a dataset is published under CC BY, anyone who gets a copy can use it and pass it on, so there's little left to sell. Open licenses also assume the data is safe to publish, and company records almost never are, even after cleanup. For internal records, open licensing is mostly useful as a way to understand the terms buyers have grown used to.

Commercial dataset licenses

A commercial license is written for one deal. The terms that matter most:

• The licensee. A named company, and whether its affiliates or customers are included.
• Permitted use. Training and evaluating the licensee's own models, nothing broader.
• Sublicensing and resale. Whether the licensee may pass the data to a lab it supplies or to anyone else.
• Exclusivity. None by default; if granted, limited by use or time and priced separately.
• Term and territory. How long the rights last and where they apply.
• Outputs. No reproduction of your records and no naming your company as the source.
• Deletion. What happens to every copy when the license ends.

Our guide to data licensing agreements goes through each clause.

Licenses on dataset platforms

If a buyer or a marketplace hosts data on a platform like Hugging Face, the dataset card carries a license field. It can name a standard license or say "other" and point to custom terms (Hugging Face). A custom commercial license should be attached in full; a label alone is how licenses end up omitted or wrong.

How to choose

Data and code you didn't create

Company datasets often contain material from somewhere else: vendor documentation, customer files, open-source code. Each keeps its own license, and you can't grant more rights than you hold. Public code datasets show the care this takes; The Stack was built only from permissively licensed code and published with a governance plan (Kocetkov et al., arXiv, November 2022). Leave third-party material out, or list it separately so the buyer knows which parts you're licensing. If you're licensing code, our guide to selling source code covers open-source components.

Common questions

Does a CC BY license allow AI training?

CC BY allows reuse for any purpose, including commercial use, as long as the creator is credited. Creative Commons also points out that its licenses can't forbid a use that copyright exceptions already allow, and that training AI may be such a use in many cases.

Can we charge for a dataset that's under an open license?

You can charge for delivering a copy, but anyone who receives it under an open license may pass it on, often for free. If you want to be paid for access, keep the data under a commercial license instead of publishing it openly.

What license should we use for a dataset we're selling to an AI company?

A custom commercial license, usually the buyer's draft reviewed by your lawyer. It should limit use to training and evaluating the buyer's own models, say whether the buyer may pass the data on, and set the term, payment and deletion terms.

Does the license cover the model trained on the data?

Only if it says so. Some open licenses, like CDLA-Permissive-2.0, explicitly put no restrictions on results. A commercial license can restrict how the model is used and what it may reproduce, but it can't make the buyer remove the data from a model that has already been trained.