AI & Copyright / Training Data Rights

Current landscape

The New York Times' case against OpenAI (and, in related actions, Microsoft) remains active as of early 2026, consolidated into multidistrict litigation in the Southern District of New York under Judge Sidney Stein, currently in contested discovery over deleted datasets and privilege claims. The Authors Guild's consolidated class action against OpenAI is likewise active and in discovery. Getty Images' case against Stability AI split into two: in the UK, Stability prevailed at trial in November 2025 after Getty abandoned its primary copyright-infringement claim mid-trial, though Getty's appeal on the surviving secondary claim has been granted and is pending; in the US, a separate Getty action is active, with a motion to dismiss heard in February 2026 and case-management conferences running through at least November 2026. Notably, Getty simultaneously announced a licensing deal with OpenAI in June 2026. Suing one AI company over training while licensing to another is a clear signal that large rights-holders are already building a two-track strategy (license where you can get paid, litigate where you can't), which is exactly the dynamic that leaves individual creators out unless a framework is built around them specifically.

The U.S. Copyright Office's own "Copyright and Artificial Intelligence, Part 3: Generative AI Training" report considered and declined to endorse a pure compulsory licensing regime, concluding the potential harms (loss of rights-holder consent and control) outweigh the benefits; most creator advocacy groups agree, generally preferring opt-in licensing over a government-set mandatory rate. On the legislative side, a bipartisan TRAIN Act would let creators access records of whether and how their work was used in training; a narrower AI Music Transparency Act (watermarking/disclosure-focused) has stalled in committee since late 2025; and a bipartisan bill from Reps. Dean and Moran, introduced January 2026, aims to protect creators from unauthorized AI training. None of this has become law as of mid-2026. Meanwhile, reporting on existing AI-training licensing deals between labels and AI companies indicates individual artists and songwriters typically receive nothing directly. The deals are struck at the label/publisher level, with no obligation to pass payment down to the people whose work trained the model.

Proposal mechanics

A collective-licensing clearinghouse — structurally similar to ASCAP/BMI for music performance rights, or the mechanical-license system for cover recordings — would collect payment from commercial AI developers based on training-corpus composition and distribute it to registered rights-holders, proportional to use where that's traceable and via a statistical sampling/estimation approach where it isn't. That estimation step is an unresolved technical challenge, named honestly rather than assumed away: verifying exactly what a frontier model trained on, at the granularity needed to pay individual creators fairly, is not a solved problem today. The TRAIN Act's proposed disclosure mechanism is the prerequisite infrastructure this depends on. You cannot get paid for a use you have no way to prove happened.

Trade-offs and opposing views

Opt-in licensing — which both the Copyright Office and most creator groups prefer over a flat compulsory rate — means some rights-holders will hold out entirely, leaving gaps in training corpora and legal uncertainty for models already trained before any license existed. This issue does not resolve the retroactivity question (a look-back window vs. prospective-only application); it's flagged here as an open legislative choice, not an oversight.

Two objections deserve a direct answer. AI labs and some economists argue any mandatory payment framework raises training costs, slows development, and — most damagingly — entrenches whichever incumbents can already afford large licensing deals (the Getty-OpenAI deal is precisely the kind of consolidation this concern points to). The party's answer is Proposal 5's safe harbor for smaller and open developers, not abandoning the payment principle. Rights-holder maximalists, on the other end, want a strict consent-required regime with no default training use at all. The party's answer: the Copyright Office itself found that a maximalist compulsory-consent approach carries its own costs, and the collective licensing marketplace model preserves more creator agency than a flat mandated rate while still solving the problem that individual, one-by-one negotiation cannot work at internet scale. That's exactly why almost no individual creator sees payment today even when a label or publisher does.