If you ask me where I learned the word “tree,” I cannot give you a neat citation. Countless people around me used it when I was a child until it stuck. Ask me what happened last Saturday in the AFL, and if I jump online, check a match report, then reply, I can point you straight to the source. That difference captures how AI uses data for training versus during inference.
For AI models, training is like childhood language exposure. Data is absorbed into numeric weights and biases, and you usually cannot map any one output back to a specific training document. Those edge cases where models regurgitate chunks aside, attribution at training time is not practical. Inference is different. If the model pulls in today’s news articles at question time, it can cite, attribute and link back. That pattern is often called retrieval augmented generation (RAG).
The opportunity: inference time licensing
This is where I see a huge opportunity for both content providers and AI model factories. Traditional bulk data deals with lump sum fees for training feel dated. Usage based licensing at inference time, i.e. when the model produces an output, lets rights holders monetise each qualified use, while users see clear attribution and links.
We are already seeing the shape of this future in high profile publisher agreements, including by Axel Springer, The Financial Times and News Corp, that combine training rights with real time use and visible attribution to original content when the model produces an output.
From recent conversations I had with some of the largest copyright holders in the world, my impression is that training deals are familiar, while inference time licensing is newer. The interest is real, the terms are still fragmented, and the leading examples above serve as reference points more than a complete playbook. That is normal for a market that is moving from bulk supply toward metered access.
What I find particularly exciting: Inference time licensing will not stop at text. As AI native user interfaces evolve almost week by week, we are seeing generative tools slip into everyday products for images, video, music and design. Imagine asking for a presentation slide and the AI pulls in a licensed graphic template with clear credit, or generating a social clip that carries an embedded credential showing which broadcaster’s footage it drew on. Standards like C2PA make this technically possible by attaching verifiable provenance to media so attribution can travel with the file. The opportunity here is that licensing moves from a hidden back-end deal to something the user actually sees in real time - a link, a label, a credential - turning provenance into both a trust signal and a revenue stream. [1]
We are also beginning to see the rise of startups that specialise in licensing at inference time. These businesses treat retrieval augmented generation not just as a technical method, but as a commercial opportunity, creating the infrastructure for usage based licensing of text, images and beyond.
I'm a strong believer in fair attribution and compensation of copyright owners. We are actively exploring inference time licensing for our own Matilda model series.
What's next?
What comes next in Australia and beyond looks clear enough. Product experiences will surface sources by default, licences will price actual use, and creators will see value return through transparent reporting. Training will not disappear, it will sit alongside inference time access with different terms. If you publish content, there is now a credible path to be discovered by AI, cited in product and paid when your work helps answer a user’s question.
