‘TRAIN Act’ Proposes Rightsholder Access to AI Training Data

Should generative AI developers be required to disclose the protected materials upon which their models trained? At least one member of Congress believes so, and the lawmaker has introduced the “TRAIN Act” as a result.

Short for the “Transparency and Responsibility for Artificial Intelligence Networks Act,” that straightforward bill comes from Senator Peter Welch (D-VT).

Spanning six pages, the TRAIN Act, in keeping with its title, would establish a subpoena-focused process through which rightsholders could determine “which of their copyrighted works have been used in the training of” AI models.

Any rightsholder who “has a subjective good faith belief that the model” at hand “used some or all of 1 or more” of their copyrighted works during training would be able to pursue a subpoena, per the bill itself.

To do so, each relevant party would “request the clerk of any United States district court to issue a subpoena to” an AI developer, seeking “records sufficient to identify with certainty, the copyrighted works” actually incorporated into training, the TRAIN Act proceeds.

Unsurprisingly, given that a number of generative AI developers have allegedly trained their products on protected music without permission, industry companies and organizations are supporting the TRAIN Act. That includes the RIAA, the Recording Academy, and the American Federation of Musicians.

Notwithstanding this backing, however, time will tell whether the measure picks up congressional momentum and becomes law. As things stand, different AI bills – chief among them the NO FAKES Act, which, in a nutshell, would establish federal voice and likeness protections – have made seemingly little legislative progress despite multiple industry pushes.

Then there’s the all-important “fair use” question. Logic suggests that ingesting troves of human creations via increasingly advanced AI models (which developers acknowledge rely heavily on protected materials) without rightsholder permission or compensation is anything but fair.

Nevertheless, AI players are adamant that this very practice does, in fact, constitute fair use; related legal battles are underway.

To state the obvious, if generative systems can lawfully train on whatever content they want, permission or no permission, related subpoenas will be of comparatively little import for rightsholders, who’ve already identified alleged infringement based on AI outputs.

Meanwhile, as litigation and legislation plod along (even the EU’s already-passed AI Act will take some time to ramp up, and training-related questions remain), artificial intelligence is continuing to evolve by leaps and bounds.

Music creation, voice replication, and video generation represent some of the many areas where companies have made noteworthy announcements to this point in November.