IMPEL and IMPF published a joint generative AI licensing framework on September 30, setting out seven principles. The central demand is that the song be valued “at least equally with the recording.” IMPEL is the London-based collective licensing agency for independent publishers’ digital rights; IMPF is the Brussels-based trade body that says it advocates for more than 300 companies. Both say any model that treats publishing as a minor share of value from the outset is unacceptable.
The framework covers revenue allocation between master and publishing rights, payment for past use, training, outputs and future exploitation, and transparency. It warns that limited pilots, closed environments or specific product models must not set broad precedents. The groups are running three attribution-technology pilots. The statement comes almost four months after the NMPA’s industry-wide deal with Udio, which the NMPA said was the first to value songs and recordings equally.
The framework turns an argument about splits into an argument about evidence: no allocation can be enforced unless both sides read the same record of use.
I learned this running revenue across a dozen countries: every partner has to read the same record, or the split falls apart. A split is only as good as the record every party reads.
Sony Music has told the Financial Times it has filed takedown notices for more than 260,000 AI-generated deepfake tracks, nearly double the roughly 135,000 it reported in March 2026 and up from 75,000 in March 2025. The recordings imitate the voices or likenesses of artists including Harry Styles, Adele, Britney Spears, Queen and Michael Jackson.
Sony says it expects platforms to act quickly and prevent the same content reappearing, calling the effort “an uphill struggle.” Deezer found 90,000 AI-generated tracks reaching on-demand services daily as of July.
The Operator’s Take: The number measures rights-holder effort, not harm: takedowns scale with headcount while uploads scale with compute.
Build · Detection & Telemetry Track notices, re-uploads and time-to-removal per platform, and make stay-down terms a requirement in your next platform negotiation.
Australia’s public broadcasters, the ABC and SBS, told Parliament’s Joint Select Committee on Artificial Intelligence that the news bargaining incentive should cover AI companies. The ABC called the current system “anti-competitive” and warned of cannibalization of news.
Anthropic accepts the ban on a text and data mining exception but asked for a “narrow form of conditional approval” to train on content unless rights holders block crawlers with robots.txt. SBS replied that robots.txt mechanisms are “routinely ignored or bypassed” and the onus should sit on AI companies. Anthropic and OpenAI executives are due at Sydney hearings the week of October 5.
The Operator’s Take: I learned this opening a whole city’s data to the public access: That access works only when the rules of use are shared, public and enforceable. An opt-out that depends on a signal the user can ignore is none of those.
Advocacy · Industry Positioning In any market weighing an opt-out model, back submissions that ask for an enforceable consent standard with audit rights rather than a crawler signal.
At AdExchanger’s Programmatic I/O in New York on Monday, September 28, panelists said insurers are becoming a new player in AI governance, writing AI into policy coverage with their own views on how much human oversight is needed. A Gartner survey in April found about 60% of companies using AI will hit unexpected charges or exceed budgets for lack of human oversight, and 56% deployed AI tools without clear policies.
The panel said AI workflows must be fully auditable: which model was active when a mistake occurred, what data was generated, and whether a human or a machine erred.
The Operator’s Take: I learned this proving local-news commitments to a regulator: the evidence has to be designed in from the start. An insurer will ask for the same log a licensor does.
Build · Detection & Telemetry Add model identifier, data inputs and human-or-machine attribution to every agent action log now, and ask your insurer what evidence it will require.