Violeta Arnaiz Medina. Director of Intellectual Property, AI and Software.
According to the RAE (Spanish Royal Academy Dictionary), “anthropic” (from the Greek antropos, “man”) means “produced or modified by human activity”. And it was indeed human activity that produced the millions of books that Anthropic used as training material for its Claude family of models, which is why it was sued by the authors of several of such books. Anthropic’s business, which generates more than $1 billion in annual revenue, revolves precisely around this technology, which allows users and the machine to communicate in natural language and the development of which requires vast amounts of text.
Some of these books, which were key to the tool’s success, were legally acquired by Anthropic and then scanned for conversion into a format that the systems could analyse. The judgment handed down by Judge Alsup in June of this year, initially interpreted as a victory for Anthropic, ruled that the activity of training Claude with these books fell within the fair use exception, as the use was considered “transformative” in the sense that there was a very large gap between the reproduced material and how it was displayed or reflected in the final result. However, other books used by Anthropic were not acquired legally, but downloaded from “pirate” websites. And here comes the part that tarnished the much-vaunted victory: the judge ruled that training a model with books that had not been accessed legally did not meet the requirements for fair use and should be declared infringing.
In July, after the judgment was announced, this part of the case became a class action lawsuit, allowing others affected to join it. The number of authors of pirated books has now reached half a million, which gives an idea of the economic significance of the case and the risk Anthropic is taking in this proceeding. Anthropic has made its move and, in a conservative play—attempting to avoid an unfavourable court ruling that would also serve as a precedent for the ruling in similar cases—has preferred to reach an agreement in which it controls the terms of the equation: what it pays authors per book used—$3,000—and the requirements that must be met in order for them to be compensated—among others, having their work registered with the US Copyright Office. The total compensation could be around $1.5 billion. However, despite the existence of a text already approved by the litigating parties, Judge Alsup is not convinced and has put the agreement on hold until he obtains more information, which the parties have already been asked to provide.
Whatever happens in this case, the concern of LLM developers about the financial and legal consequences of lawsuits filed against them for the use of copyrighted material is evident. The rulings handed down in these proceedings will define the rules of the game. A game that, so far, has only been beneficial to one of the parties: the developers, who defend themselves by saying that had they not used the “better to ask for forgiveness than permission” technique, they would not have been technologically capable of getting this far.
At this point, rights holders are beginning to take action. Just a couple of weeks ago, the Swedish management entity STIM (the equivalent of the Spanish SGAE or General Society of Authors and Publishers) announced the launch of a licence that allows the legal use of music recordings for training AI models in exchange for remuneration for rights holders—a model with certain similarities to the extended collective licensing project that emerged some time ago in Spain. The STIM licence is combined with a technology called Sureel, which allows any AI-generated additions to a human creation to be tracked, making the process traceable. This system, which aims to become a standard, allows creators to earn income proportional to the intensity with which AI uses their works, and developers to train their models in a secure and transparent manner. Norway has also just reached an economic agreement with the country’s newspaper publishers to use their content in a generative AI system that the government is developing. And in the United States, companies such as the Copyright Clearance Centre (CCC) – cited in the lawsuit against Anthropic as an example of a legal option for carrying out these activities without infringing rights – have long existed, allowing the use of content protected under licence.
For many, myself included, a business model that rewards all agents who contribute to the value chain is the only sustainable solution in the medium/long term. There are precedents: if we look back, we will remember the digital transformation process of the music industry, and how, while technology and the law played cat and mouse, parasitic business models emerged and profited from music without paying for it. These models disappeared once the system was readjusted and an attractive legal offering was created for users and providers.
AI is here to stay, and the decision on how we integrate it into our productive environment and what role we give to human creation in the value chain it generates is up to us—humans—and it is up to us now. The future will be Anthropic, or it will not be.
