Legal Limits of Training Commercial AI on Copyrighted Text

The legal boundary between "innovation" and "infringement" in artificial intelligence has reached a definitive turning point. On September 29, 2026, the U.S. Third Circuit Court of Appeals affirmed a pivotal ruling against ROSS Intelligence, determining that its unauthorized use of Westlaw’s copyrighted legal summaries to train a competing AI model did not constitute fair use. This decision serves as a stark warning: the era of consequence-free scraping is over, and the market value of verified, licensable datasets has never been higher.

The Four-Factor Test: Why Scraped Data is a Liability

For years, AI developers relied on the "transformative use" defense under Section 107 of the U.S. Copyright Act (copyright.gov). However, the ROSS Intelligence ruling clarifies that simply feeding copyrighted text into a neural network to output similar functional data is not inherently transformative. For a commercial AI model to claim fair use, it must generally provide a new purpose or character that does not supersede the objects of the original work.

For data buyers, using unlicensed datasets now carries a high risk of "contributory infringement." If the training data was obtained by bypassing paywalls or violating Terms of Service (ToS), the resulting model may be legally tainted. This has led to a surge in demand for premium dataset catalogues where provenance and usage rights are explicitly cleared for commercial machine learning.

The EU AI Act and the TDM Exception

While U.S. law focuses on fair use, the European Union has codified a more structured approach through the EU AI Act. Article 53 of the Act (eur-lex.europa.eu) mandates that providers of general-purpose AI models must put in place a policy to respect Union copyright law, specifically the Text and Data Mining (TDM) exception.

Under this framework, commercial entities can only mine data if the rightsholder has not explicitly "opted out" using machine-readable means (such as robots.txt or metadata). For organizations sitting on vast archives of proprietary text, this means that active technical management of your IP is now a prerequisite for monetization. You must ensure your data is properly indexed and protected to maintain its value in the licensing market.

The Economic Reality: Disclosed vs. Estimated Deal Values

The shift toward legal compliance is driving massive capital flows into the data economy. Major publishers are no longer litigating for pennies; they are signing multi-year licensing agreements. For instance, News Corp disclosed a landmark deal with OpenAI valued at over $250 million over five years (wsj.com), while industry analysts at Reuters estimated the Axel Springer-OpenAI partnership to be worth between $25 million and $30 million annually (reuters.com).

These figures demonstrate that "clean" data is a premium asset class. Data owners should focus on three pillars to prepare for monetization:

• Provenance Logging: Maintaining a clear chain of custody for every record.
• Rights Mapping: Identifying which portions of the data are proprietary vs. third-party.
• Compliance Auditing: Ensuring the data meets the standards outlined in our guide on what you can legally sell under GDPR and AI regulations.

Decision Framework for AI Data Acquisition

If you are an AI team or an integrator, your acquisition strategy must shift from volume to verifiability. A checklist for your next data deal should include:

1. Does the seller provide a warranty against third-party copyright claims?
2. Is the data gathered in compliance with the EU AI Act’s TDM opt-out requirements?
3. Does the license grant "perpetual" rights to the trained weights, even if the data license expires?

Failing to answer these questions can lead to "model deletion" orders, where a court forces a company to destroy a model trained on infringing data—a catastrophic loss of R&D investment.

What this means for you

The 2026 legal landscape has effectively ended the "Wild West" of data scraping. For data owners, your proprietary text is now a high-yield asset, provided you have the legal documentation to back it up. For data buyers, the cost of licensing is significantly lower than the cost of a copyright injunction. Whether you are looking to list a specialized corpus or acquire high-fidelity training sets, d-nvest provides the intelligence and the marketplace to execute these deals with total legal certainty.