Attorneys for OpenAI and The New York Times are arguing over training data in a federal courtroom in New York. Not in the abstract—about particular articles, particular archival material, and particular works that the Times claims were included into sizable language models without authorization or compensation. The Times filed its complaint in late 2023, and since then, the discovery process has turned into a battlefield of its own. The Times is suing OpenAI for allegedly withholding internal tools that might search training data for content protected by copyright. The plaintiffs think there is proof of infringement, and they think the defendant has been tardy to provide it. This is the kind of legal detail that sounds arcane until you grasp what it means.
This is just one aspect of the generative AI industry’s growing legal dilemma. The statutory damages clause of copyright law permits plaintiffs to recover up to $150,000 per intentionally infringed work, although it is difficult to determine the exact financial risk in the event that courts finally find against the top AI developers. The math gets worrisome when the works in question are measured in the millions. The industry’s potential liability is easily estimated to be in the hundreds of billions of dollars, and in worst-case scenarios, some legal analysts have put the figure even higher.
Right now, the legal basis is truly precarious. The claim that training AI on data constitutes a transformative use that is eligible for fair use protection—a legal theory that permits restricted use of copyrighted content without authorization for purposes like commentary, education, and creative transformation—has been approved by several federal judges. Some judges have been less persuaded, especially when it comes to the amount of material covered and the training businesses’ commercial orientation. These models’ outputs immediately compete with the works they were trained on, which is precisely the scenario that copyright law is intended to handle. It remains to be seen if that argument is successful.
An additional layer of exposure is introduced by the issue of pirated data. The inclusion of information from shadow libraries and unlicensed repositories in the LAION-5B dataset, which is frequently used to train image-generating machines, has drawn significant criticism. Compared to the overall web-scraping issue, judges have taken a far more stringent stance on this. When it was discovered that Anthropic’s training data contained millions of books that had been downloaded illegally, the company was exposed to substantial financial risk.
When the source is a well-known piracy library, it is more difficult to maintain the claim that a corporation was unaware of the provenance of every piece of data in a training corpus. The legal coalition chasing generative AI developers now includes over 400 local newspapers in 33 states, which has significantly changed the nature of the battle.
Large media conglomerates and digital corporations are no longer the only parties involved in this conflict. It involves small regional newsrooms with little funding to continue lawsuit against well-funded internet giants since they rely only on the value of their original reporting. Their participation has altered public perceptions and could eventually alter the legal approach.
In response, the industry has mostly argued for fair usage while discreetly negotiating licensing agreements with specific publications; this strategy indirectly recognizes some danger while attempting to mitigate it. AI developers may encounter something far more disruptive than settlements if the courts ultimately reject the fair use argument at scale: obligations to retrain models without infringing data, which would require starting anew in ways that would take years and cost billions of dollars.
