Content licensing for AI inference: what platforms should know

Content licensing for AI inference is at the heart of our new partnership with Cashmere, and it differs from licensing content to train a model. Platforms are better served by treating the two as separate rights with separate terms. In licensing, inference covers any use in which a model reads source material, often retrieved at the time of the question, to generate something new for a user. An AI research tool that summarizes a set of articles is doing inference. So is a chatbot that grounds its reply in a reporter's coverage.

Public debate about AI and content has focused largely on training: whether models may learn from copyrighted work and what that should cost. Training matters. But many AI products depend on something else, which is access to current and reliable source material at the moment a question is asked. That use needs its own license, with its own terms.

This article explains the difference, why inference changes the economics of licensing and what a platform should ask for before it signs. For a wider view of how licensing works, see our content licensing Knowledge Hub.

Three different AI use cases call for three sets of terms

AI uses content in three broad ways, and each is a different activity.

• Direct use is when AI shows content back to a person in substantially its original form, such as reproducing, narrating or translating it.
• Inference is when AI reads content in order to generate something new, such as a grounded chat answer, a retrieval-augmented research report, a summary, or a recommendation.
• Training is when content is used to improve the model itself.

Cashmere, a platform that helps premium publishers license their content to AI applications, draws this line clearly in its executive brief, Pricing Content in the Age of AI, which focuses on direct use and inference. We find the distinction useful because terms written for one of these do not automatically fit the others. One product can also involve more than one use case. An AI reading assistant, for example, may display an article to a user and then answer questions about it.

Inference changes the economics because the model reads the source

A person can read an article only so many times. A model can draw on the same article for an enormous number of answers, and no person ever needs to open the original. Cashmere's brief describes this as a different economic model altogether, and we agree. A license built around a one-time sale or an undifferentiated page view does not describe how the content is actually used.

Usage varies a great deal. A consumer chatbot answering a simple question may touch only a small piece of text. A deep research workflow may synthesize hundreds of sources and depend heavily on their quality. The value of the product matters too. Cashmere's brief points out that a general chatbot, a university tutor, an enterprise research assistant, or a clinical support tool can consume the same text and produce very different economic outcomes. Take one article on a clinical trial. A consumer chatbot might draw a single sentence from it. A research workflow built for pharmaceutical teams might base an entire analysis on it. The content is the same, but the dependence on it and the value created are not.

Content licensing for inference should take the particular use case into account.

Scraping gives a platform access to pages but not to terms

Scraping can retrieve what is openly available on the internet. It cannot give a platform a license. It generally comes with no negotiated usage terms, no reporting back to the source, and often no access to content behind a paywall. In addition, the content pipeline can break when websites change layouts or block the bots that attempt to scrape them.

To produce a quick product demo, scraping may be enough. For a product that customers pay for, the question changes. It is no longer only whether the content can be fetched, but whether the platform can say where an answer came from and under what terms the content was used. Customers and compliance teams tend to ask exactly that.

Licensed content can answer it. The permissions are written down, access is set by a contract, and the license can require a certain amount of usage reporting. Publishers stay inside the value chain instead of outside it.

An inference license should answer seven questions

Price often dominates licensing talks, but the terms around it can be more important in practice. Cashmere's brief makes the same point: a well-priced agreement with poor terms can cause more trouble than a somewhat mispriced one if the terms are well thought out. Before signing an inference license, a platform should be able to answer these:

1. Permitted use. What may the platform do with the content (ground answers, summarize, show excerpts) and what is off limits, such as training a model? Without this, a license meant for answering questions can be read to cover more than either side intended.
2. Attribution. How are sources credited to the end user in an answer? Attribution lets readers check an answer and gives publishers credit for the work behind it.
3. Usage reporting. What does the platform report back to the content supplier, and how often? Reporting is how the content supplier, and the publishers behind it, learn how much the content is really used.
4. Retention. May the platform store the content, and for how long? Stored copies can outlive the agreement that allowed them.
5. Sublicensing. May the platform pass the content to its own customers or to other applications? Passing it on changes the scale of use.
6. Audit. How can the supplier verify that the terms are being followed? License agreements that do not enable some form of auditing rely on trust alone.
7. Enforcement. What happens if the terms of a license agreement are not observed? Clear remedies keep a small breach from becoming a dispute.

Each of these issues can be addressed in different ways, depending on the needs of the platform and the content supplier. The point is that every one should have an answer in writing before the content goes into production.

Cashmere chose Newstex to add breadth to its catalog

A platform building an inference product often needs both depth and breadth in its content sources. Depth comes from scholarly, professional and reference material. Breadth comes from timely reporting, commentary and specialist coverage. These often come from different suppliers with different terms.

Cashmere, a new Newstex customer, shows how that plays out. Cashmere is building its catalog on academic and professional publishing, with publishers licensing their content through its infrastructure for AI inference. As more AI applications are coming to Cashmere for licensed sources, they are asking for more variety, including timely reporting, commentary and a long tail of specialty coverage that scholarship does not reach.

That is the gap Newstex fills. Newstex is a curated content licensing service that supplies licensed news, commentary, and specialist coverage to information platforms and AI-enabled products. Under the new partnership, Newstex supplies that coverage to Cashmere, which can make it available to AI applications through its license-gated infrastructure. Cashmere's infrastructure governs access for inference, checks licenses on every query, and keeps content out of model training unless the rights holder permits it.

For platforms, that means depth and breadth through one licensed channel. For Newstex publishers, it means an additional licensed route to AI applications.

Newstex describes the principles that fair AI content licensing requires in Trust, capability, reward: A framework for fair AI content licensing. In Pricing Content in the Age of AI, Cashmere describes a path forward that is based on these principles.

Building an AI product that needs licensed news, commentary and specialty coverage? Request a proposal and the Newstex team will follow up with information and, if you're ready, a demo.

‍