The first wave of artificial intelligence (AI) regulation focused on what goes into models. California’s Training Data Transparency Act, which went into effect Jan. 1, requires developers to disclose their training data, including sources, collection periods and whether data includes copyrighted or personal information. The European Union (EU) published a similar mandatory disclosure template under the AI Act, requiring providers to identify data sources, modalities and processing methods. Both rules assume that transparency about inputs will produce accountability for outputs.
The argument gaining ground now is that what a model produces matters more than how it was built, and that existing laws are better equipped to address measurable harms than to govern training pipelines.
Google published a 21-page AI governance paper on Thursday (June 25) that makes that case directly. “This approach would address outputs, not inputs,” the paper stated, “looking to prevent and mitigate specific harms rather than micromanaging the science behind these new tools.” Written by Kent Walker, Google’s president of global affairs, the paper arrives as regulators in California, Colorado and the EU are moving in different directions, with some focusing on training data disclosure and others on real-world outcomes.
On copyright specifically, the paper argued that training on publicly available web data “is a transformative, non-expressive use — like an art student taking inspiration from walking through a gallery — that should remain protected under fair use in the U.S.” Copyright concerns, the paper added, should focus on whether a specific image or piece of text actually copies an existing work, regardless of how it was created.
For publishers who don’t want their content used in training, Google pointed to machine-readable opt-out controls like robots.txt as the appropriate mechanism, with paid agreements covering specialized or non-public content. On liability more broadly, the paper continued: “People on the left and the right agree that if something is illegal to do without AI, it’s illegal to do with AI — we don’t need to reinvent the wheel.”
Output-Based Regulation Gains Ground as Input Rules Multiply
The paper’s output-first argument runs directly against the direction California and the EU have taken. Early disclosures by OpenAI and Anthropic under California’s statute acknowledged that training data may include copyrighted material but provided limited detail, PYMNTS reported, reflecting both practical constraints and unresolved legal questions around large-scale data scraping.
Colorado’s revised approach is the closest to Google’s position. PYMNTS reported that Colorado revised its AI Act in May to shift regulatory focus away from the technology itself and toward the real-world outcomes AI systems produce. Businesses operating in Colorado must now determine where AI materially influences a consequential decision, and, if that influence leads to an adverse outcome, notify the affected consumer. The law’s overarching theme, per an analysis by law firm Mintz, is governance of outcomes, not inputs.
Split on Where AI Liability Should Start
The judicial record mirrors the regulatory divide. U.S. District Judge William Alsup in San Francisco called AI training “quintessentially transformative” and said copyright law “seeks to advance original works of authorship, not to protect authors against competition,” PYMNTS reported.
U.S. District Judge Vince Chhabria, also in San Francisco, reached a different conclusion two days later, warning that widespread AI training could undermine the economic incentives that drive human creative work, according to PYMNTS. More cases involving Anthropic, Google and Stability AI are pending in 2026.
Publishers are pushing back on the opt-out model Google recommends. Digital Content Next sent a cease-and-desist letter to the Common Crawl Foundation, reported by Press Gazette. The letter argued that copyright law “is not an opt-out regime,” a direct challenge to Google’s position.
Daryl Lim, H. Laddie Montague Jr. Chair in Law at Penn State Dickinson Law School, told PYMNTS that only a handful of firms can train frontier models at scale, since those firms simultaneously control compute, data, cloud infrastructure, and distribution.
“When you train frontier models, you need to ingest vast repositories of works that may include copyrighted works,” Lim said. That concentration is the structural concern an output-based framework alone does not reach.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
