Only 4 in 10 AI Licensing Deals Still Include Training Rights

David Martin Riveros · July 31, 2026 · 4 min read

Rob Kelly keeps the most complete public ledger of AI content licensing deals I know of, a running spreadsheet at Media & the Machine of every publicly announced agreement since 2023. His July 10 analysis covers 94 deals, and buried in it is a number I have not seen quoted anywhere else: only about 4 in 10 of those deals include training rights. Through 2023 and the first half of 2024, training was the standard clause, the thing the deal existed to sell. A publisher handed over its archive, a lab absorbed it into a model, and everyone argued about the price of the past. Most deals signed now leave that clause out.

His June update holds the series that shows where the money went instead. Deals granting attribution or live access, the right to fetch a publisher's content at answer time and cite it, went from 2 in 2023 to 11 in 2024 to 18 in 2025, with 34 projected for this year. Kelly projects 36 total deals in 2026. Put those two figures side by side: nearly the entire deal flow of 2026 is live access. The market did not cool, it changed products.

Why the training clause is disappearing

Kelly lists five drivers, and they compress into two stories. The buyer's story: frontier models stopped improving much from one more pile of general web text, so the marginal archive is worth less every quarter, while the products people actually use answer questions by retrieving current pages at the moment you ask. A lab building that needs standing access and permission to cite, not a copy of 2019. The seller's story: publishers learned the asymmetry the hard way. A feed can be repriced at renewal or switched off in a dispute. Content baked into model weights is gone for good at whatever price you first accepted, and the training clause is also where the copyright litigation concentrates. Meta's line in the ledger makes the shift concrete: of its 11 tracked deals, Kelly notes, only one, News Corp, mentions training at all.

Nobody buys the same file twice

I run a company that collects web data on a schedule, and this is the least surprising chart I have seen all year, because recurring delivery has been the entire business the whole time. A consulting client does not buy retail prices, it buys retail prices re-collected every week, because the analysis it sells is about what changed. A travel client watches rental rates that the booking sites requote within a day. In years of doing this, no customer has ever paid me twice for an old file. When someone asks for history, it is to backtest or audit something they built on the live feed. The commercial weight of a dataset sits entirely on the next delivery.

The labs arrived at the same billing model from the other direction. An archive sale prices the past once. A feed prices the present at every renewal, and either side can reopen the number, which is exactly why sellers prefer it and why buyers accepted it once training stopped being the point. If you hold content and are still waiting for your training check, the 4-in-10 figure is the polite version of the news. The window when archives cleared at a premium ran from 2023 to about the middle of 2024. What still commands recurring money is whatever you publish next, fetched fresh, cited, and repriceable.

One caveat from inside the ledger. Kelly counts public deals only, and his own estimate is that every announced deal may sit on top of 50 to 100 private ones. He reckons private, training-focused arrangements still clear 75 to 100 million dollars a year. So the archive is not dead, it trades quietly and at a discount, out of sight of the market that sets reference prices. But price discovery happens in public, and the public market has decided. In 2023, AI paid for what the web had already written. In 2026, it pays to be let back in tomorrow morning.