Collective licensing and opt-out infrastructure for training data

Rights-holders — authors, artists, musicians, journalists, photographers, publishers, code authors — aggregate their individual data-and-copyright interests into collective structures (rights-collecting societies, opt-out registries, industry-wide licensing consortia, standard-form training-data-use terms) that constrain the AI vendor's training-data supply on collective terms rather than on the vendor's single-actor terms. The infrastructure ranges from opt-out registries (a machine-readable "do-not-train" claim asserted at the domain, feed, or work level) through provenance-and-consent registries (Content Authenticity Initiative, C2PA) to full collective-licensing agencies on the ASCAP / BMI model (a single license from the collective covers the full member repertoire, with revenue distribution back to members). The strategy converts a diffuse individual-rights problem — every creator would need to sue individually to defend their own work — into a collective-bargaining position with the vendor at the training-data supply layer.

An actor chooses this strategy because the training-data supply chain is the AI model's structural chokepoint: a large-scale model cannot be trained without access to a large corpus, and the corpus is composed of individual works whose rights-holders can, in aggregate, refuse or price that access. The strategy runs at the supply layer rather than at the deployment layer, which means it constrains the vendor's whole downstream product regardless of which harm the model produces. It also plays the standard music-industry precedent: the ASCAP / BMI collective-licensing regime resolved a comparable individual-rights collective-action problem in the 1910s-1940s by aggregating songwriters' rights into a single license the broadcaster had to negotiate against. The AI industry sits in exactly the pre-ASCAP position — training on individual works without a collective counterparty — and the strategy is to build that counterparty.

It trades scope for aggregation. A collective licensing regime only reaches works whose rights-holders have opted in (or, in a compulsory-licensing regime, whose statute requires the licensing) — the vast public-domain and unclaimed corpus remains available. The strategy is vulnerable to the sui generis legal argument that AI training is transformative use or otherwise exempt from copyright, an argument the vendor bar has pushed hard in every jurisdiction; a court ruling either way reshapes what the collective can license and on what terms. And the collective itself is a governance problem: rights-collecting societies have historically distributed revenue unevenly toward established repertoires and against new / marginalised creators, replicating industry hierarchies inside the movement-side infrastructure.

Ecology

Distinct from creator-class collective bargaining on generative AI. That strategy operates at the labour-contract layer — creator unions bargaining with employer studios and publishers over generative-AI use in the workplaces where creators are employed. This strategy operates at the rights-in-work layer — rights-holders bargaining with AI vendors over the works themselves as training data. The two strategies often live inside the same creator-sector organisations (the Authors Guild pursues both; the Writers Guild has pushed both), but they address different vendor-relationships (employer vs. AI-vendor) and different bargaining objects (employment terms vs. training-data licensing).

Distinct from class-action litigation against private-sector AI harms. Class actions are backward-looking litigation for damages on training that has already happened, resolved through the courts on a per-case basis. Collective licensing is forward-looking infrastructure that establishes the licensing regime the next round of training will operate inside. The two strategies complement — a class action can produce a settlement that includes a licensing commitment, and a collective licensing regime is often what a class action settles into — but the strategies work at different times in the vendor's product cycle.

Distinct from distributing user-side technical countermeasures. User-side countermeasures (Glaze, Nightshade, robots.txt at scale) are technical tools individual rights-holders deploy against training; collective licensing is institutional infrastructure the aggregate rights-holder deploys against the vendor. The two are complementary: the countermeasures raise the cost of unauthorised training, which strengthens the licensing collective's bargaining position by making the alternative to a license more expensive.

Fed by empirical audit and expose. Documented evidence that a vendor trained on a specific corpus (books, articles, code) is the standing evidentiary base for a rights-holder collective's licensing demand; the audit converts a suspicion of unauthorised training into a citable finding the collective can bring to the negotiating table.

Coordinated with class-action litigation against private-sector AI harms at the litigation-and-licensing pipeline. The class action supplies the leverage; the collective licensing regime supplies the settlement architecture the leverage lands in. A well-organised creator-sector coalition runs the litigation arm to establish the vendor's liability and the licensing arm to convert that liability into an ongoing revenue-and-consent regime.