A new US bill wants to unmask AI's training data

AI companies have long treated their training data as trade secrets. When pressed on what their models actually learned from, the standard answer is usually a vague gesture toward publicly available internet data.

That free ride might be coming to an end.

On Tuesday, California Democratic Representative Adam Schiff introduced the Generative AI Copyright Disclosure Act. The bill targets the heart of generative AI, aiming to force companies to reveal the copyrighted works powering their systems.

A 30-day countdown for AI datasets

The proposed law would require AI companies to file a complete list of copyrighted works included in their training datasets with the Register of Copyrights. Crucially, companies would have to submit this documentation at least 30 days before releasing any new generative tool that makes text, images, music, or video.

Fail to file, and companies face financial penalties. Given that modern datasets contain billions of text lines, countless images, or millions of hours of audio and video, producing that paper trail is a massive logistical challenge.

Unsurprisingly, the entertainment industry is fully on board. First reported by Billboard, the bill quickly scored endorsements from major creative heavyweights, including the Recording Industry Association of America (RIAA), the Directors Guild of America (DGA), the Professional Photographers of America, and SAG-AFTRA.

Duncan Crabtree-Ireland, national executive director of SAG-AFTRA, pointed out that everything generated by AI ultimately originates from human creativity. The bill also comes right after more than 200 high-profile musical artists signed an open letter demanding protection against AI tools that could replace them.

The crumbling wall of fair use

The legislation lands right in the middle of a legal war zone between tech giants and creators.

OpenAI is already drowning in copyright lawsuits, facing claims from authors like Sarah Silverman and institutions like The New York Times. According to The Washington Post, the startup hired an army of top-tier attorneys to handle over a dozen active cases.

So far, tech firms have leaned heavily on fair use, claiming their ingestion of copyrighted content is legal. In a filing to a UK government committee earlier this year, OpenAI explicitly stated that copyright law does not forbid training, adding that its models simply would not function without access to copyrighted material.

Schiff's bill does not actually ban companies from scraping copyrighted work. Instead, it forces them into radical transparency.

Why it matters

We think this bill takes a remarkably clever tactical angle. Instead of wading into the murky, decade-long legal swamp of redefining fair use, it simply demands procedural transparency. Right now, artists are trapped in a catch-22: they suspect their work was scraped, but they cannot prove what is inside a closed model. Forcing companies to post a receipt removes that barrier instantly.

For creators and media companies, this could completely shift the balance of power. With training registries on file, negotiating licensing deals stops being an impossible shot in the dark and becomes a straightforward auditing process.

However, the compliance burden could easily backfire against open-source developers and smaller startups. Parsing billions of data points for copyright provenance requires immense legal resources. Giants like OpenAI, Microsoft, and Google can absorb that friction, but smaller teams could be crushed by it, effectively cementing the power of the incumbents.

There is also the practical question of enforcement. It remains unclear whether the US Copyright Office even has the infrastructure to inspect Petabytes of metadata, or if the statutory fines will just become an ordinary cost of doing business for trillion-dollar tech firms. Still, the era of frictionless, consequence-free data harvesting is clearly drawing to a close.