New York's AI Transparency Bill: A Barrier to Innovation?

New York's AI Transparency Bill: A Potential Barrier to Innovation

New York’s proposed AI Transparency Bill aimed to require artificial intelligence developers to disclose the data sources used in their models. While intended to promote transparency, critics raised concerns that the bill could create significant compliance challenges, potentially stifling innovation in the AI sector. The bill, introduced by State Sen. Kristen Gonzalez, has stalled, but its implications continue to spark debate.

The bill, known as Assembly Bill 8595, or the Artificial Intelligence Transparency for Journalism Act, mandated AI developers to provide detailed records of every URL and data source used during the development of their models. This requirement was particularly targeted at "journalism providers," defined as media outlets that perform a public information function and invest significant resources in their operations. Under the bill, these publications would have the right to sue for damages or seek injunctive relief if their data was used without proper disclosure.

Technical Challenges of Compliance

Complying with the bill’s requirements would have presented significant technical difficulties. Large language models (LLMs), which are central to many AI applications, rely on vast datasets collected through web crawlers. Tracking and documenting every URL accessed during the development process is not a standard practice in the industry. Furthermore, LLMs learn by adjusting numerical values that represent connections between words, rather than storing specific URLs. Once trained, a model’s knowledge reflects broader patterns derived from the data, not individual sources.

"Tracking every URL used in model development is not only impractical but also technically infeasible," said Andrew Mayne, an AI consultant and novelist. "Large language models don’t store URLs; they learn from patterns in the data." This technical challenge raised questions about the bill’s feasibility and its potential impact on the development of AI models.

Verification and Documentation Issues

The bill also raised concerns about the verification process. Engineers often manually verify information, such as clinicians checking medical citations. Under AB 8595, companies might have been required to document every URL visited during these verification checks. However, experts pointed out that a URL alone does not prove that a model’s understanding came solely from that link. Discussions, promotional materials, and other sources could also influence a model’s output.

"A URL is just one piece of the puzzle," Mayne noted. "It doesn’t tell the whole story of how a model learns and processes information." This ambiguity highlighted the complexity of documenting the influences on AI models and the challenges of enforcing transparency measures.

Copyright and Fair Use

The bill did not clearly define what constituted a copyright violation, which could have made it easier for publishers to prove infringement. Recent legal cases have begun to define the boundaries of fair use in the context of AI. In September, AI developer Anthropic agreed to a $1.5 billion settlement with authors, setting a precedent that AI models can be trained on legally obtained works without violating copyright.

In a related case, a U.S. district judge ruled that Meta’s use of books to train its AI did not substantially harm publishers. However, in Thomson Reuters v. Ross Intelligence, a judge ruled that using proprietary content to train an AI engine was not fair use, emphasizing the originality and commercial nature of the content.

Legal Precedents and Future Implications

The bill raised important questions about the balance between transparency and innovation in AI development. While the bill stalled in the rules committee in June 2025, the tensions between developers and publishers persist. Experts suggest that similar legislation could resurface in New York or other states, highlighting the ongoing debate over AI transparency and its impact on the industry.

"Even with perfect logs, developers couldn’t trace all influences on model outputs," Mayne said. "The bill raises questions about disclosing URLs used for fact-checks or ancillary information, and its impact on the development process." As AI continues to evolve, the need for clear guidelines and balanced legislation will become increasingly important.