AI content licensing

AI content licensing is the sale, by publishers, news agencies and other owners of writing, photos or video, of the right to use their content in AI models. The buyers are the companies that build large language models and the chatbots and search tools built on them. A licence is a written permission, given in return for payment. It states what the buyer may do with the content, and for how long.

The buyer may want the content for two different uses. The first is training. A model learns from huge amounts of text, and a large archive of edited, checked articles is valuable material to learn from. The second is answering. When a chatbot or an AI search tool answers a question about today's news, it fetches current articles and writes a summary from them. This method is called retrieval-augmented generation (RAG). For this use, the publisher wants the answer to name it and link to its page.

Here is how a deal is usually made. The publisher works out what it owns: its archive, its new articles each day, and its photos. It agrees with the AI company what the content may be used for: training, answering, or both. It agrees the price. This may be a fixed yearly fee, a payment each time an article is used in an answer, or both. It then delivers the content through a regular feed, and tracks how it is used. The price is the hardest part, because nobody yet knows what one article is worth to a model. Software for AI content licensing and pricing tries to put a value on each article.

Why is this happening now? AI companies collected much of their training text with crawlers, programs that copy pages from the open web. Many publishers now block those crawlers. Some have gone to court, arguing that copying their articles without permission breaks copyright law. In India, a news agency filed such a case in the Delhi High Court in 2024. A licence gives both sides an agreed way forward. Publishers also need the new income, because they are getting fewer readers from search, as described in zero-click search.

Take, for example, a news agency. It already sells its stories to newspapers by subscription. It can also license its archive of several decades to an AI company, as one more customer.

Consider an academic publisher. It can license the papers on its journal publishing platform to a company that is training a model on scientific writing.

Say a photo agency owns millions of pictures, each with a caption. A company training an image model needs exactly that: pictures paired with descriptions.

How is AI content licensing different from syndication? In syndication, a news agency sells a story so that a newspaper can publish it to readers, under the agency's name. In AI content licensing, the content may never be shown to anyone as it is. A model learns from it, or quotes a few lines of it in an answer. So the deal must also settle how the publisher is credited, and how it is paid when nobody visits its page.