AI training licensing deals are starting to feel like yesterday\u2019s news as publishers and platforms focus on more dynamic, usage-based models.<\/p>\n
Rather than the initial training deals that formed the backbone of AI licensing partnerships between AI platforms and news publishers, recent deals have forged around different parameters: what many in the industry refer to as \u201cAI grounding.\u201d<\/p>\n
In fast-moving digital areas like AI, the terminology tends to splinter quickly. Vendors, publishers, platforms and analysts coin their own terms: for instance, \u201cgrounding,\u201d \u201ccontent inference compute\u201d, and \u201cretrieval augmented generation\u201d (RAG) are all intertwined and refer more or less to the same thing. Those who can\u2019t be bothered with jargon of any sort simply call grounding and RAG \u201cweb search.\u201d\u00a0<\/p>\n
To AI engineers, there are subtle differences between them, but for publishers, RAG\/grounding has changed how they get paid now given how the large language models (LLMs) now process information.<\/p>\n
One-time lump sum payments are out; recurring, usage-based licensing agreements are in. \u201cAs we\u2019ve moved more into RAG deals, the per-usage aspect of these pricing structures has become the preeminent piece of the pie when it comes to fees,\u201d said Aaron G. Rubin, partner in the strategic transactions and licensing group for law firm Gunderson Dettmer.\u00a0<\/p>\n
Here\u2019s a primer.\u00a0<\/p>\n
What is the difference between training versus grounding deals?<\/strong><\/p>\n In a nutshell, payment terms of grounding or \u201cRAG\u201d deals are based on how AI systems fetch live content from publishers in real time. If a person searches for an update on some recent news like, \u201cShow me an update on the meeting between Trump and Zelensky,\u201d which happened over the last week, AI engines won\u2019t have that stored in their training. \u201cTraining windows for AI engines tend to be up to six months old; they don\u2019t know anything after the training date,\u201d said Martin Alderson, co-founder of web performance consultancy Catch Metrics. That\u2019s why they use RAG to pull the information from a multitude of publishers to provide the best response to the user.\u00a0<\/p>\n That model should create opportunities for recurring licensing revenue, attributions and continued visibility. In contrast, training deals are typically one-time payments where publishers get an upfront lump sum, or have a fixed fee over years for content used to train a model. The New York Times agreed to a training deal with Amazon, to the tune of $20 million, while News Corp did similar for $50 million. Many of the agreements from the first wave of publisher-AI platform deals would have been for training.<\/p>\n Why is focus shifting to so-called grounding or RAG deals?<\/strong><\/p>\n For starters, few publishers would have been able to negotiate to the same level as the NYT and News Corp. But also because the value of training data has receded for AI platforms. For publishers like DPG Media, training deals don\u2019t warrant decent payouts, stressed Valerie de Naeyer, head of Gen AI transformation and operational excellence at DPG Media. \u201cIn terms of copyright law, publishers are not so keen on licensing content to train the model either \u2014 lots of questions on IP remain unresolved,\u201d she said. \u201cIt\u2019s possible that there is also a training component in some deals, in case of historic or less relevant content, but in case of real-time, content grounding is preferred,\u201d she added.\u00a0<\/p>\n On July 30, Gannett signed a licensing deal with Perplexity to allow it to license content from USA Today and the USA Network. As always, details on payment terms are scarce, but it\u2019s an example of a RAG\/grounding deal due to Perplexity\u2019s approach, which centers on ad revenue sharing, not training content deals.\u00a0<\/p>\n