Why real-time access is becoming the most valuable thing in content licensing

Much of the conversation between AI companies and publishers has been about training. Who scraped what, when, and whether anyone paid for it. And training was traditionally seen as a one-time transaction rather than an ongoing relationship. However, retrieval-augmented generation (RAG) is changing the game. 

RAG shifts content licensing from a one-time transaction to ongoing supply

RAG refers to the practice of having a model look up relevant information at the moment a question is asked and then basing its answer on what it finds. The now-ubiquitous AI overviews at the top of Google’s search results are one example of how this can work in practice.

Training cutoffs leave models without current facts, and RAG fills the gap 

RAG is important because all large language models have training cutoffs. Anything that happened after that date might as well have happened in another dimension. However, the model won’t necessarily know what it’s missing and this can lead to misleading answers based on out-of-date material.

The only way to solve this problem is to give the model access to current information. But without some Doctor Who-type shenanigans, that material won’t be in the dataset the model was trained on. RAG can offset this problem.

Facts decay at different rates, so freshness needs vary by content type 

Facts decay at different rates.

Some facts are forever–King Charles III will always have ascended the throne on September 8, 2022, for example. But other facts are far less fixed. Think share prices, changes in political leadership, or the emergence of new technologies. In situations like these, even a week-old corpus can be out of date. It’s little wonder therefore that Meta's CRAG benchmark found that models performed noticeably worse on questions involving complex, niche topics. 

RAG raises accuracy but depends on the quality of the corpus

RAG can help models access the kind of specialized content that will enable them to provide high-value answers. But RAG isn’t a silver bullet since even the most current data can have weaknesses. After all, a model’s output is only as good as the material it can access. 

The same CRAG evaluation is instructive here. Models answering from their own parameters alone performed poorly. Adding retrieval in a straightforward way improved things but the best industry systems still flubbed a significant share of their answers by providing incorrect information or failing to provide any answer at all. 

A strong RAG corpus will consist of authoritative material that is structured well enough that the right passage can easily be found. Content should also be as complete as possible to lessen the risk that the retriever will silently paper over gaps with something adjacent and wrong.

In this environment, vetted content is more important than ever.

RAG makes source attribution traceable, which gives publishers leverage 

There is a second consequence of RAG that publishers should not overlook.

When content is used in training, its contribution disappears into the weights. You cannot point at a model output and say that sentence came from such-and-such an article. That opacity has been a real obstacle in every licensing negotiation and ignited much of the litigation.

RAG works differently. Since the system knows which documents it pulled, it can show its work. The chain from source to answer is meant to be traceable because the architecture requires it. This makes attribution technically straightforward in a way it never was with training. It also makes unlicensed use considerably harder to defend. A platform that retrieves from a document it has no right to use is effectively tattling on itself.

For publishers, that traceability provides them with leverage.

It also raises the value of being licensed and properly documented because a platform that has to answer for its sources would rather answer with a contract than with an explanation.

Platforms running RAG need continuously refreshed, structured feeds 

The commercial implication follows from the technical one. If a platform's answer quality depends on a continuously current corpus, then it needs a continuously current supply of information. That in turn requires reliable delivery infrastructure, consistent structure, and accurate metadata. A feed that arrives in an unpredictable shape creates work the platform may eventually decide is not worth the hassle.

It also means the relevant question for a publisher has shifted. Value is no longer decided by a one-time question. Now, it has become an ongoing proposition.

This is the shift Newstex is built around: Newstex delivers licensed, provenance-documented content as structured, continuously refreshed feeds, the freshness platforms increasingly require and the recurring distribution publishers increasingly need. Think syndication could be right for you? Apply here.

‍