It’s widely understood that AI models behind chatbots such as ChatGPT, Gemini, and Claude leverage colossal databases of published content, including vast collections of books, online articles, and academic papers—essentially, a significant portion of the internet. Many authors have, unknowingly and without granting permission, contributed to the very AI technologies now perceived as a threat to their professions. This raises an immediate question of legality, doesn’t it?
However, the legal situation is far from straightforward.
“This entire legal and technological domain is characterized by immense activity and complexity,” noted Cathy Gellis, an attorney specializing in intellectual property, copyright, and technology, in an interview with TechCrunch. She added, “There are deep emotions on both sides regarding these developments.”
In a landmark decision last year, Judge William Alsup mandated that Anthropic pay a substantial $1.5 billion copyright settlement to a collective of writers whose material contributed to the training of the company’s AI models. While initially appearing as a win for authors, Judge Alsup’s ruling actually affirmed the legality of Anthropic’s AI training process. The penalty against Anthropic was imposed specifically for obtaining these copyrighted works from illicit online ‘shadow libraries,’ rather than for the act of training itself.
The judge elaborated, stating that “Anthropic’s LLMs, much like an aspiring writer studying literature, trained on these works not to merely copy or replace them, but to significantly diverge and generate something novel.” He drew a parallel between an LLM’s assimilation of vast textual data and a human writer’s process of learning from existing literature.
Gellis perceives this ruling as largely beneficial for AI enterprises, questioning the impact of a $1.5 billion fine on a company expected to reach approximately $200 billion in yearly revenue by 2028.
Gellis commented, “It’s generally positive news for AI training that the judge viewed the activity as analogous to reading copyrighted material rather than outright copying it.” She clarified that “Copyright law focuses on the act of copying, not on the mere use, experience, consumption, or reading of a work.”
The Copyright Act, unchanged since 1976, forces judges to apply decades-old principles to novel legal challenges posed by AI, issues that could significantly influence the industry’s future trajectory.
“There is widespread concern currently due to the inconsistent application of the law, a direct result of this unresolved question,” stated Jason Henderson, Senior Attorney and Founder of the IP & Media Practice at JWL International, to TechCrunch. He emphasized, “Stakeholders are aware of the extensive data AI models are trained on, and the legal framework has not yet adequately addressed this specific issue.”
Such legal inquiries frequently revolve around the doctrine of fair use—specifically, whether the utilization of a copyrighted work is sufficiently “transformative” to be deemed legally acceptable.
Fair use serves as an exception within copyright law, permitting the use of copyrighted content without direct authorization, thereby safeguarding the capacity to critique, parody, educate, and otherwise build upon existing works. When assessing fair use, judges typically weigh factors such as the work’s purpose and character, the quantity of the original work utilized, and its potential effect on the market.
“Copyright inherently aims to safeguard and expand the market,” Henderson observed. He added that “court rulings in AI-related cases vary widely in their rationale. A prevailing trend indicates that if the intent behind training an AI on another’s property is to directly compete, courts are likely to disapprove… Conversely, if the use does not pose direct competition, courts are more inclined to find it permissible.”
Henderson refers to a lawsuit where media and technology giant Thomson Reuters pursued legal action against research firm Ross Intelligence, alleging that Ross copied its content to develop a rival AI-powered legal research platform.
Last year, Judge Stephanos Bibas concluded, “Ross’s usage lacks transformative quality as it fails to demonstrate a ‘further purpose or different character’ compared to Thomson Reuters’s.” In this instance, Judge Bibas ruled against fair use, deeming it impermissible to train on Reuters’ material to create a directly competitive platform. Although authors might contend that chatbots generating synthetic books constitute direct competition, this specific argument has yet to succeed in litigation.
Regarding the intersection of AI and copyright, Gellis advises distinguishing between two distinct concepts: the legal framework surrounding AI training and the considerations for copyrighting content produced by AI, as these are approached quite differently.
A significant ruling in Thaler v. Perlmutter established that entirely AI-generated works are not eligible for copyright protection. This decision introduces complex challenges, raising questions about how to conclusively determine the extent of AI involvement in a work’s creation and what percentage of it might have been AI-assisted or generated.
Gellis drew an analogy: “When you compose a novel using Microsoft Word and utilize its spell-check feature, we readily accept that Word does not claim ownership of your novel.” She concluded, “AI is compelling us to revisit numerous assumptions that have largely gone unchallenged.”
With most AI companies currently entangled in ongoing lawsuits regarding these matters, a clear and definitive resolution to these complex problems is not expected in the immediate future.
Gellis observed, “The initial legal challenges are proving influential, though their impact could be reversed by divergent rulings from other courts; subsequent stages of litigation will ultimately determine which precedents prevail.” She emphasized, “Nevertheless, these ongoing decisions are actively shaping the entire landscape, and it would be unwise for AI companies to disregard them.”