Insights on licensed, attributed AI access to authoritative content.
The sticker price on an AI license says nothing about margin. Two deals at the identical headline rate can diverge by €170k in cumulative contribution once preparation costs, run costs and royalty shares are priced in. Here is the formula to calculate the floor before negotiating.
Francois-Xavier Bioul
5 min read
A discovery order in a books-and-AI-training case, and a fresh amicus brief from AAP, News/Media Alliance and STM in the same litigation, both treat the AI licensing market as more than a backdrop to fair use. What that means for how companies should document licensing negotiations from now on.
Article 50(4) exempts human-reviewed AI text from disclosure, but the exemption protects the version a person actually reviewed. A faithful AI translation stays inside that protection; an AI summary, rewrite or SEO pass that changes structure or meaning does not, and the published version needs its own review.
On August 27, 2026, the Fifth Circuit held in Emmerich v. Particle Media that URLs are not categorically excluded from copyright management information under the DMCA — but most permalinks will not qualify as drafted. The same ruling replaced the server test with a transmit requirement for framing. A four-factor audit for publisher permalinks before any AI or aggregator deal.
The CMA has given publishers a world-first opt-out from Google's AI features. wikiHow's complaint against OpenAI shows what a control is worth when nobody logs it. Six objects, ranked by who actually holds the evidence — and why crawler logs come first.
ALA's new AI guidance gives libraries a 10-item vendor review checklist covering logging, training use, retention and audit rights. For vendors and publishers selling into libraries, most items are still answered with policy, not proof.
Scientific publishers are tightening AI disclosure policies, but disclosure alone cannot reconstruct how claims, citations, figures or review decisions were produced. STM publishers need a second layer: evidence records that make AI-assisted editorial workflows observable — starting where disclosure fails first, with citations.
Anthropic's latest policy position highlights a larger problem for publishers and rights holders: static AI disclosure rules can become obsolete as models change. Future-proof transparency should focus less on technical architecture and more on durable, rights-aware evidence of content usage.
The Third Circuit's review of Thomson Reuters v. Ross Intelligence is not only about whether AI training can qualify as fair use. It exposes a sharper commercial question: should content that improves an authorized product be governed like content used to build a competing substitute?
Anthropic tied a $21.6bn Australian investment to copyright clarity, arguing the "long tail" of small rights holders can't be licensed one deal at a time. Canberra refused a shortcut anyway. That leaves the operational question open — and some organisations are already answering it.
Folha de S.Paulo settled its OpenAI lawsuit in May 2026 — no financial terms disclosed. Cloudflare's new Pay Per Use model and Microsoft's Publisher Content Marketplace both promise usage-based payment instead. Neither answers who audits the number.
The first AI licensing deals are becoming market reference points while their terms remain confidential. The real benchmark question is not the headline price. It is who controls the usage data behind it — and whether publishers can verify the buyer’s number.
Mind if we use cookies?
We use a privacy-friendly analytics cookie to see how Citations Logic is used. You're in control — accept or reject anytime.
