Courts Split Over AI Training Fair Use Rulings

U.S. District Judge William Alsup ruled in June 2025 that training AI models on copyrighted books purchased and digitized for that purpose can qualify as fair use, calling it "quintessentially transformative" (PYMNTS). Two days later, U.S. District Judge Vince Chhabria, also in San Francisco, reached the opposite conclusion in a related case, warning that widespread AI training could undermine the economic incentives behind human creative work (PYMNTS; Reuters). More cases involving Anthropic, Google and Stability AI are pending in 2026. Against that backdrop, Google published a 21-page AI governance paper on June 25, 2026, arguing that "this approach would address outputs, not inputs... rather than micromanaging the science behind these new tools," and that training on publicly available web data is "a transformative, non-expressive use... that should remain protected under fair use in the U.S." Google recommends machine-readable opt-out tools like robots.txt for publishers who object. Digital Content Next rejected that framing in a cease-and-desist letter to the Common Crawl Foundation, arguing copyright law "is not an opt-out regime" (Press Gazette).

1. June 24, 2025

   Judge William Alsup rules that training AI models on purchased, digitized copyrighted books can be fair use.

2. June 26, 2025

   Judge Vince Chhabria reaches the opposite conclusion in a related San Francisco case involving AI training.

3. January 1, 2026

   California's Training Data Transparency Act takes effect, requiring AI developers to disclose training-data sources.

4. June 25, 2026

   Google publishes a white paper arguing AI regulation should target outputs rather than training inputs.

5. July 1, 2026

   PYMNTS reports that Digital Content Next has sent Common Crawl a cease-and-desist letter over web-scraping opt-outs.

California and the EU have taken an input-focused approach, requiring disclosure of training-data sources, collection periods, and copyright status. Colorado moved the other way, revising its AI Act in May 2026 to focus on real-world consequential decisions rather than the underlying technology, per a Mintz analysis cited by PYMNTS. Google's paper explicitly sides with the output-based camp, and separately proposes a federally-overseen, industry-funded body (which it calls FARO) to set safety standards for frontier models.

The Alsup/Chhabria split has not been resolved by appeal or settled into binding precedent, so dataset provenance, retention practices, and documented transformation remain the concrete factors courts have weighed. Teams should not treat Google's fair-use argument or the output-based regulatory framing as settled law; California and Colorado's disclosure obligations already apply regardless of how the fair-use question is eventually resolved, and licensing exposure differs by jurisdiction.

Whether appellate courts or additional district rulings in the pending Anthropic, Google, and Stability AI cases narrow or widen the Alsup/Chhabria split; how regulators respond to Google's FARO proposal and output-based framing; and whether the Digital Content Next-Common Crawl dispute produces a test case on whether robots.txt-style opt-outs are legally sufficient.

The pattern here is a familiar one in emerging-tech regulation: dominant players that already operate at scale tend to favor output-based, harms-focused rules that are lighter on upfront compliance, while critics warn that structural concentration in compute, data, and distribution needs input-side scrutiny to be addressed at all. Neither framework has won yet, and practitioners should expect both fronts to keep moving in parallel.