I have been working in content licensing for more than 20 years. But what is currently described as “AI licensing” is something I still struggle to understand.
No buyers
If you ask AI model developers about their actual budget for content licensing, you usually get a very clear answer. Guess how much that budget is. The economics and business logic of foundation model development leave very little room for meaningful payments per individual work.
No sellers
As far as I can see, creators and many rightsholders have not been presented with concrete licensing offers at a broader scale. And even if they were, expectations on the rightsholder side may not align with the reality of what AI companies are prepared to pay.
No licensing agreements
To make a bold statement: most of the publicly reported agreements between AI companies and rightsholders are not "licensing agreements" in the strict legal sense. They are "access deals", "commercial partnerships", or "arrangements" designed to avoid legal disputes in the future.
Proper licensing of digital content requires the consent of the original creators and underlying rightsholders. Anyone who has ever tried to license a single song or film knows how complex that is – even when serious money is involved. Buyouts, in sectors where they are common practice, simplify negotiations. But they do not resolve the deeper issues of rights clearance, in general.
No sustainability
If payments are calculated per individual work within massive training datasets, the amounts will be minimal. Legal scholar Martin Senftleben argues that licensing input data for AI training is unlikely to be sustainable at scale. A levy or collective remuneration model – where AI systems pay into a fund that is distributed by collecting societies – may be more realistic and the direction the EU parliament is headed. But such infrastructure – focussed on the output of AI systems rather than the input for AI training – would take years to communicate and implement.
Conflicting legal positions
AI companies are unlikely to openly embrace broad “AI licensing” if it weakens their litigation strategy. In the US, companies rely on fair use. It is difficult to argue in court that training is lawful under fair use while at the same time building a comprehensive licensing market for the same activity (and not only for content behind paywalls).
No real need
As long as companies rely on fair use in the US and on Article 4 DSM in Europe – and as long as publicly available content remains accessible at scale because the opt-out option is not widely exercised by creators and rightsholders – the economic incentive to pay broadly for training data remains limited.
A token payment remains a token. The argument that higher “data quality” will drive large-scale licensing markets often sounds more convincing on the seller side than on the buyer side.
Marketplaces may facilitate certain access deals. But the idea of a large-scale, sustainable licensing market for AI training data remains questionable.
Also, there may be an unintended side effect: a licensing model under the current legal regimes tends to reward rightsholders who technically restrict access to their content. Communities that have deliberately chosen open licenses are now starting to question that decision because the situation feels structurally unfair, see the discussion around Creative Commons preference signals.
The real debate should not focus on headlines that oversimplify the issue. It should focus on whether the economics and the legal framework genuinely support what is currently being described as “AI licensing.”